{"meta":{"title":"World Models \u00d7 LLMs \u2014 RAG Cortex","generated":"corpus-08","note":"Two hemispheres \u2014 World Models (sensing) | LLMs (reasoning) \u2014 joined at the corpus callosum (Liminal). Every paper in the corpus is a point; theme hubs + key papers are planets; colored nodes are typed knowledge-graph entities (topics, domains, methods, terms). Edges are semantic kNN / theme membership plus typed KG relations (introduces / uses / evaluated_on / about).","counts":{"concepts":40,"papers":521,"entities":150,"nodes":711,"edges":3541,"libraryPoints":12045,"total":12756}},"silhouette":[[0.5,0.0868],[0.5104,0.0761],[0.5212,0.0682],[0.5322,0.064],[0.543,0.0638],[0.5536,0.0664],[0.5639,0.0701],[0.5743,0.073],[0.585,0.0736],[0.5965,0.0717],[0.6086,0.068],[0.6211,0.0641],[0.6334,0.0623],[0.6448,0.0641],[0.6547,0.0704],[0.6628,0.0811],[0.6693,0.0946],[0.6749,0.109],[0.6806,0.1223],[0.6871,0.1327],[0.6953,0.1394],[0.7055,0.1426],[0.7174,0.1431],[0.7308,0.1421],[0.7449,0.1408],[0.7592,0.14],[0.7733,0.1402],[0.7872,0.1416],[0.8007,0.144],[0.8139,0.1474],[0.8265,0.152],[0.8379,0.1582],[0.8477,0.1663],[0.8553,0.1765],[0.8608,0.1886],[0.8644,0.2019],[0.8671,0.2157],[0.8699,0.229],[0.874,0.2411],[0.8799,0.2518],[0.8877,0.2614],[0.8964,0.2706],[0.9045,0.2803],[0.9106,0.2913],[0.9134,0.3038],[0.9127,0.3178],[0.9094,0.3325],[0.9051,0.3471],[0.9021,0.3607],[0.9026,0.3729],[0.908,0.3835],[0.9185,0.3929],[0.9331,0.4019],[0.9497,0.4112],[0.9657,0.4214],[0.9789,0.4328],[0.9877,0.4454],[0.9915,0.4588],[0.9911,0.4726],[0.9878,0.4864],[0.983,0.5],[0.9783,0.5134],[0.9743,0.5265],[0.9711,0.5395],[0.9682,0.5525],[0.9651,0.5653],[0.9612,0.5779],[0.9565,0.5902],[0.9511,0.6022],[0.9457,0.6141],[0.9407,0.6259],[0.9362,0.6377],[0.932,0.6496],[0.9277,0.6614],[0.9226,0.6729],[0.9165,0.6839],[0.9096,0.6944],[0.9024,0.7046],[0.8963,0.7152],[0.8922,0.727],[0.8909,0.7406],[0.8923,0.7562],[0.8955,0.7738],[0.8989,0.7922],[0.9004,0.8101],[0.8983,0.8258],[0.8915,0.8379],[0.8799,0.8459],[0.8645,0.8499],[0.847,0.851],[0.8294,0.8511],[0.8133,0.852],[0.7999,0.855],[0.7892,0.8609],[0.7806,0.8693],[0.773,0.8792],[0.7651,0.889],[0.7562,0.8973],[0.7457,0.9033],[0.734,0.907],[0.7215,0.909],[0.7091,0.9106],[0.6974,0.9131],[0.6868,0.9174],[0.6771,0.924],[0.6681,0.9327],[0.6594,0.9427],[0.6506,0.9533],[0.6413,0.9637],[0.6316,0.9735],[0.6213,0.9825],[0.6105,0.9905],[0.5992,0.9972],[0.5873,1.0019],[0.5748,1.0034],[0.5618,1.0008],[0.5487,0.9934],[0.5356,0.9816],[0.523,0.9668],[0.5111,0.9514],[0.5,0.938],[0.4895,0.9292],[0.479,0.9265],[0.4683,0.93],[0.4568,0.9383],[0.4446,0.9489],[0.4318,0.9589],[0.419,0.9659],[0.4066,0.9682],[0.3951,0.9658],[0.3845,0.9596],[0.3745,0.9517],[0.3645,0.9446],[0.3538,0.9402],[0.3417,0.9395],[0.3281,0.9425],[0.313,0.9477],[0.2972,0.9533],[0.2815,0.9571],[0.267,0.9575],[0.2543,0.9537],[0.2437,0.9458],[0.2352,0.9346],[0.2284,0.9213],[0.2225,0.9072],[0.217,0.8931],[0.2118,0.8794],[0.2065,0.8663],[0.2013,0.8537],[0.196,0.8415],[0.1904,0.83],[0.1841,0.8196],[0.1765,0.8105],[0.1672,0.8029],[0.1562,0.7968],[0.1438,0.7914],[0.1309,0.7859],[0.1186,0.7794],[0.1082,0.7712],[0.1001,0.7612],[0.0945,0.7496],[0.0903,0.7371],[0.0861,0.7248],[0.0805,0.7133],[0.0724,0.703],[0.0613,0.6937],[0.0482,0.6849],[0.0349,0.6756],[0.0235,0.665],[0.0163,0.6527],[0.0146,0.6386],[0.0188,0.6231],[0.0277,0.607],[0.0393,0.591],[0.0511,0.5758],[0.0606,0.5617],[0.0664,0.5486],[0.0679,0.5362],[0.0659,0.5243],[0.0616,0.5122],[0.0565,0.5],[0.0521,0.4875],[0.0491,0.4748],[0.0475,0.462],[0.0469,0.4492],[0.0468,0.4364],[0.0465,0.4234],[0.0461,0.4103],[0.0457,0.3971],[0.046,0.3838],[0.0473,0.3707],[0.0499,0.3579],[0.0534,0.3453],[0.0575,0.333],[0.0615,0.3206],[0.0651,0.308],[0.0685,0.2952],[0.0724,0.2826],[0.0779,0.2708],[0.0861,0.2604],[0.0976,0.2524],[0.1125,0.2469],[0.1297,0.2437],[0.1476,0.2418],[0.1642,0.2399],[0.1777,0.2363],[0.1869,0.2297],[0.1918,0.2194],[0.1934,0.2057],[0.1932,0.1897],[0.1933,0.1731],[0.1954,0.1578],[0.2004,0.1453],[0.2085,0.1362],[0.219,0.1302],[0.2309,0.1261],[0.2427,0.1226],[0.2537,0.1181],[0.2635,0.1118],[0.2722,0.1038],[0.2805,0.0947],[0.289,0.0857],[0.2984,0.0782],[0.309,0.0731],[0.3208,0.0709],[0.3334,0.0713],[0.3465,0.0736],[0.3595,0.0772],[0.3724,0.0814],[0.3849,0.0859],[0.3972,0.0909],[0.4091,0.0965],[0.4208,0.1026],[0.432,0.1089],[0.4427,0.1146],[0.4529,0.1183],[0.4624,0.1189],[0.4716,0.1157],[0.4807,0.1085],[0.4901,0.0983]],"regions":{"t0":{"hemi":"wm","label":"Embodied AI & Robotics","lobe":"frontal","ax":0.2933,"ay":0.6291,"color":"#9a4f4f","count":859,"hub":"concept_t0"},"t1":{"hemi":"llm","label":"LLM Training & Scaling","lobe":"parietal","ax":0.7534,"ay":0.6691,"color":"#b07267","count":437,"hub":"concept_t1"},"t2":{"hemi":"llm","label":"Consciousness and Predictive Models","lobe":"temporal","ax":0.61,"ay":0.7324,"color":"#b07d67","count":14,"hub":"concept_t2"},"t3":{"hemi":"wm","label":"Flow Matching for Generative Models","lobe":"occipital","ax":0.4187,"ay":0.2044,"color":"#9a714f","count":169,"hub":"concept_t3"},"t4":{"hemi":"wm","label":"Energy System Optimization","lobe":"frontal","ax":0.2487,"ay":0.1749,"color":"#9a7c4f","count":32,"hub":"concept_t4"},"t5":{"hemi":"llm","label":"Predictive Cognition and AI","lobe":"parietal","ax":0.584,"ay":0.5992,"color":"#b09e67","count":264,"hub":"concept_t5"},"t6":{"hemi":"llm","label":"Causal World Models","lobe":"temporal","ax":0.5454,"ay":0.5393,"color":"#b0a967","count":230,"hub":"concept_t6"},"t7":{"hemi":"llm","label":"Embodied AI & Physical Reasoning","lobe":"occipital","ax":0.552,"ay":0.6076,"color":"#adb067","count":173,"hub":"concept_t7"},"t8":{"hemi":"llm","label":"Robustness and Security in LLMs","lobe":"frontal","ax":0.7775,"ay":0.6882,"color":"#a2b067","count":207,"hub":"concept_t8"},"t9":{"hemi":"llm","label":"Speculative Decoding in LLMs","lobe":"parietal","ax":0.8543,"ay":0.787,"color":"#97b067","count":183,"hub":"concept_t9"},"t10":{"hemi":"llm","label":"LLM Agent Communication","lobe":"temporal","ax":0.7514,"ay":0.5566,"color":"#8cb067","count":2,"hub":"concept_t10"},"t11":{"hemi":"llm","label":"Audio Language Models","lobe":"occipital","ax":0.8442,"ay":0.2759,"color":"#81b067","count":294,"hub":"concept_t11"},"t12":{"hemi":"wm","label":"Spatial Reasoning in VLMs","lobe":"frontal","ax":0.5595,"ay":0.309,"color":"#5e9a4f","count":472,"hub":"concept_t12"},"t13":{"hemi":"llm","label":"Medical Vision-Language Models","lobe":"parietal","ax":0.6587,"ay":0.1813,"color":"#6bb067","count":203,"hub":"concept_t13"},"t14":{"hemi":"llm","label":"Active Inference in AI","lobe":"temporal","ax":0.4992,"ay":0.7216,"color":"#67b06e","count":119,"hub":"concept_t14"},"t15":{"hemi":"wm","label":"3D Video World Models","lobe":"occipital","ax":0.425,"ay":0.5849,"color":"#4f9a62","count":435,"hub":"concept_t15"},"t16":{"hemi":"llm","label":"Sparse and Interpretable LLMs","lobe":"frontal","ax":0.7316,"ay":0.4194,"color":"#67b084","count":489,"hub":"concept_t16"},"t17":{"hemi":"wm","label":"World Modeling & Embodied AI","lobe":"parietal","ax":0.3435,"ay":0.6357,"color":"#4f9a79","count":695,"hub":"concept_t17"},"t18":{"hemi":"llm","label":"KV Cache Compression","lobe":"temporal","ax":0.8946,"ay":0.4416,"color":"#67b09a","count":206,"hub":"concept_t18"},"t19":{"hemi":"llm","label":"LLM Agent Planning","lobe":"occipital","ax":0.6359,"ay":0.8465,"color":"#67b0a5","count":485,"hub":"concept_t19"},"t20":{"hemi":"llm","label":"Predictive Coding in Deep Learning","lobe":"frontal","ax":0.5633,"ay":0.6071,"color":"#67b0b0","count":343,"hub":"concept_t20"},"t21":{"hemi":"wm","label":"High Energy Physics Phenomenology","lobe":"parietal","ax":0.11,"ay":0.785,"color":"#4f8f9a","count":77,"hub":"concept_t21"},"t22":{"hemi":"wm","label":"Multi-Agent Reinforcement Learning","lobe":"temporal","ax":0.4288,"ay":0.5922,"color":"#4f849a","count":211,"hub":"concept_t22"},"t23":{"hemi":"wm","label":"Vision-Language Models","lobe":"occipital","ax":0.4115,"ay":0.3172,"color":"#4f799a","count":501,"hub":"concept_t23"},"t24":{"hemi":"llm","label":"LLM Inference Optimization","lobe":"frontal","ax":0.8952,"ay":0.5848,"color":"#6784b0","count":271,"hub":"concept_t24"},"t25":{"hemi":"llm","label":"Free Energy in Brain Theory","lobe":"parietal","ax":0.5859,"ay":0.8215,"color":"#6779b0","count":106,"hub":"concept_t25"},"t26":{"hemi":"wm","label":"Model-Based Reinforcement Learning","lobe":"temporal","ax":0.2226,"ay":0.3587,"color":"#4f579a","count":642,"hub":"concept_t26"},"t27":{"hemi":"llm","label":"Efficient Reasoning in LLMs","lobe":"occipital","ax":0.8189,"ay":0.8,"color":"#6b67b0","count":374,"hub":"concept_t27"},"t28":{"hemi":"llm","label":"Memory and Cognitive Maps","lobe":"frontal","ax":0.5739,"ay":0.4498,"color":"#7667b0","count":347,"hub":"concept_t28"},"t29":{"hemi":"wm","label":"Video Generation & World Models","lobe":"parietal","ax":0.4229,"ay":0.5568,"color":"#6a4f9a","count":513,"hub":"concept_t29"},"t30":{"hemi":"llm","label":"Multimodal Vision-Language Models","lobe":"temporal","ax":0.7861,"ay":0.2237,"color":"#8c67b0","count":735,"hub":"concept_t30"},"t31":{"hemi":"llm","label":"Retrieval-Augmented LLMs","lobe":"occipital","ax":0.8641,"ay":0.6378,"color":"#9767b0","count":179,"hub":"concept_t31"},"t32":{"hemi":"wm","label":"Diffusion Model Advancements","lobe":"frontal","ax":0.3917,"ay":0.2142,"color":"#8b4f9a","count":422,"hub":"concept_t32"},"t33":{"hemi":"llm","label":"LLM Evaluation Benchmarks","lobe":"parietal","ax":0.8448,"ay":0.6006,"color":"#ad67b0","count":304,"hub":"concept_t33"},"t34":{"hemi":"wm","label":"Diffusion Weather Forecasting","lobe":"temporal","ax":0.342,"ay":0.1901,"color":"#9a4f93","count":59,"hub":"concept_t34"},"t35":{"hemi":"llm","label":"LLM Agent Evaluation","lobe":"occipital","ax":0.6494,"ay":0.9284,"color":"#b0679e","count":419,"hub":"concept_t35"},"t36":{"hemi":"llm","label":"Cascading Dynamics in Networks","lobe":"frontal","ax":0.6419,"ay":0.6378,"color":"#b06793","count":87,"hub":"concept_t36"},"t37":{"hemi":"llm","label":"Quantization Techniques for LLMs","lobe":"parietal","ax":0.8731,"ay":0.3767,"color":"#b06788","count":163,"hub":"concept_t37"},"t38":{"hemi":"wm","label":"World Models for Physical AI","lobe":"temporal","ax":0.3988,"ay":0.7802,"color":"#9a4f66","count":296,"hub":"concept_t38"},"t39":{"hemi":"wm","label":"Agentic Wireless Intelligence","lobe":"occipital","ax":0.2865,"ay":0.4519,"color":"#9a4f5b","count":105,"hub":"concept_t39"}},"nodes":[{"id":"concept_t0","kind":"concept","n":1,"label":"Embodied AI & Robotics","authors":"","year":null,"url":"","region":"t0","hemi":"wm","lobe":"frontal","color":"#9a4f4f","indegree":23,"alsoIn":[],"x":0.2933,"y":0.6291,"text":"Research exploring vision-language-action models and reinforcement learning for embodied manipulation and robotics."},{"id":"concept_t1","kind":"concept","n":2,"label":"LLM Training & Scaling","authors":"","year":null,"url":"","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":28,"alsoIn":[],"x":0.7534,"y":0.6691,"text":"Research focused on improving large language model training, scaling, and pre-training strategies."},{"id":"concept_t2","kind":"concept","n":3,"label":"Consciousness and Predictive Models","authors":"","year":null,"url":"","region":"t2","hemi":"llm","lobe":"temporal","color":"#b07d67","indegree":8,"alsoIn":[],"x":0.61,"y":0.7324,"text":"Exploring consciousness in AI and the brain through predictive processing and theoretical neuroscience."},{"id":"concept_t3","kind":"concept","n":4,"label":"Flow Matching for Generative Models","authors":"","year":null,"url":"","region":"t3","hemi":"wm","lobe":"occipital","color":"#9a714f","indegree":9,"alsoIn":[],"x":0.4187,"y":0.2044,"text":"This cluster explores flow matching as a framework for generative modeling, integrating concepts from optimal transport, energy-based models, and reinforcement learning."},{"id":"concept_t4","kind":"concept","n":5,"label":"Energy System Optimization","authors":"","year":null,"url":"","region":"t4","hemi":"wm","lobe":"frontal","color":"#9a7c4f","indegree":8,"alsoIn":[],"x":0.2487,"y":0.1749,"text":"Research focused on optimizing integrated energy systems using reinforcement learning, planning, and data-driven methods."},{"id":"concept_t5","kind":"concept","n":6,"label":"Predictive Cognition and AI","authors":"","year":null,"url":"","region":"t5","hemi":"llm","lobe":"parietal","color":"#b09e67","indegree":16,"alsoIn":[],"x":0.584,"y":0.5992,"text":"This cluster explores how predictive models and cognitive principles inform both neuroscience and AI, particularly in memory, learning, and physical reasoning."},{"id":"concept_t6","kind":"concept","n":7,"label":"Causal World Models","authors":"","year":null,"url":"","region":"t6","hemi":"llm","lobe":"temporal","color":"#b0a967","indegree":24,"alsoIn":[],"x":0.5454,"y":0.5393,"text":"This cluster explores methods for learning causal and interpretable world models to enhance reasoning, planning, and generalization in AI systems."},{"id":"concept_t7","kind":"concept","n":8,"label":"Embodied AI & Physical Reasoning","authors":"","year":null,"url":"","region":"t7","hemi":"llm","lobe":"occipital","color":"#adb067","indegree":12,"alsoIn":[],"x":0.552,"y":0.6076,"text":"This cluster explores the integration of embodied cognition, physical reasoning, and world models in AI systems and robotics."},{"id":"concept_t8","kind":"concept","n":9,"label":"Robustness and Security in LLMs","authors":"","year":null,"url":"","region":"t8","hemi":"llm","lobe":"frontal","color":"#a2b067","indegree":9,"alsoIn":[],"x":0.7775,"y":0.6882,"text":"This cluster explores adversarial robustness, security threats, and defense mechanisms for large language models and agents."},{"id":"concept_t9","kind":"concept","n":10,"label":"Speculative Decoding in LLMs","authors":"","year":null,"url":"","region":"t9","hemi":"llm","lobe":"parietal","color":"#97b067","indegree":17,"alsoIn":[],"x":0.8543,"y":0.787,"text":"A cluster of papers exploring efficient large language model inference through speculative decoding techniques."},{"id":"concept_t10","kind":"concept","n":11,"label":"LLM Agent Communication","authors":"","year":null,"url":"","region":"t10","hemi":"llm","lobe":"temporal","color":"#8cb067","indegree":2,"alsoIn":[],"x":0.7514,"y":0.5566,"text":"A cluster exploring LLM agent capabilities, including multilingual communication, safety, and transparent training methodologies."},{"id":"concept_t11","kind":"concept","n":12,"label":"Audio Language Models","authors":"","year":null,"url":"","region":"t11","hemi":"llm","lobe":"occipital","color":"#81b067","indegree":9,"alsoIn":[],"x":0.8442,"y":0.2759,"text":"Research focused on developing general-purpose audio language models capable of understanding and generating speech, music, and sound."},{"id":"concept_t12","kind":"concept","n":13,"label":"Spatial Reasoning in VLMs","authors":"","year":null,"url":"","region":"t12","hemi":"wm","lobe":"frontal","color":"#5e9a4f","indegree":15,"alsoIn":[],"x":0.5595,"y":0.309,"text":"This cluster explores spatial reasoning capabilities of vision-language models in multimodal and embodied AI settings."},{"id":"concept_t13","kind":"concept","n":14,"label":"Medical Vision-Language Models","authors":"","year":null,"url":"","region":"t13","hemi":"llm","lobe":"parietal","color":"#6bb067","indegree":8,"alsoIn":[],"x":0.6587,"y":0.1813,"text":"Research focused on adapting and scaling vision-language models for medical imaging and clinical applications."},{"id":"concept_t14","kind":"concept","n":15,"label":"Active Inference in AI","authors":"","year":null,"url":"","region":"t14","hemi":"llm","lobe":"temporal","color":"#67b06e","indegree":19,"alsoIn":[],"x":0.4992,"y":0.7216,"text":"This cluster explores active inference as a framework for planning, perception, and decision-making in embodied AI systems."},{"id":"concept_t15","kind":"concept","n":16,"label":"3D Video World Models","authors":"","year":null,"url":"","region":"t15","hemi":"wm","lobe":"occipital","color":"#4f9a62","indegree":15,"alsoIn":[],"x":0.425,"y":0.5849,"text":"Research focused on generating and understanding 3D scenes through video diffusion models with geometric and temporal reasoning."},{"id":"concept_t16","kind":"concept","n":17,"label":"Sparse and Interpretable LLMs","authors":"","year":null,"url":"","region":"t16","hemi":"llm","lobe":"frontal","color":"#67b084","indegree":12,"alsoIn":[],"x":0.7316,"y":0.4194,"text":"Research focused on improving the efficiency, interpretability, and safety of large language models through sparse architectures and interpretable features."},{"id":"concept_t17","kind":"concept","n":18,"label":"World Modeling & Embodied AI","authors":"","year":null,"url":"","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":76,"alsoIn":[],"x":0.3435,"y":0.6357,"text":"Research focused on building and utilizing world models for embodied AI systems to enable physical reasoning, planning, and control through video generation and diffusion models."},{"id":"concept_t18","kind":"concept","n":19,"label":"KV Cache Compression","authors":"","year":null,"url":"","region":"t18","hemi":"llm","lobe":"temporal","color":"#67b09a","indegree":12,"alsoIn":[],"x":0.8946,"y":0.4416,"text":"Research focused on optimizing LLM inference through efficient KV cache compression techniques."},{"id":"concept_t19","kind":"concept","n":20,"label":"LLM Agent Planning","authors":"","year":null,"url":"","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":38,"alsoIn":[],"x":0.6359,"y":0.8465,"text":"This cluster explores how large language models can be used as world models and memory systems to enable planning and decision-making in autonomous agents."},{"id":"concept_t20","kind":"concept","n":21,"label":"Predictive Coding in Deep Learning","authors":"","year":null,"url":"","region":"t20","hemi":"llm","lobe":"frontal","color":"#67b0b0","indegree":14,"alsoIn":[],"x":0.5633,"y":0.6071,"text":"This cluster explores predictive coding as a biologically inspired framework for deep learning, focusing on its potential to replace or enhance backpropagation."},{"id":"concept_t21","kind":"concept","n":22,"label":"High Energy Physics Phenomenology","authors":"","year":null,"url":"","region":"t21","hemi":"wm","lobe":"parietal","color":"#4f8f9a","indegree":0,"alsoIn":[],"x":0.11,"y":0.785,"text":"This cluster explores physical phenomena and parameter extraction in high-energy collisions and cosmic surveys."},{"id":"concept_t22","kind":"concept","n":23,"label":"Multi-Agent Reinforcement Learning","authors":"","year":null,"url":"","region":"t22","hemi":"wm","lobe":"temporal","color":"#4f849a","indegree":16,"alsoIn":[],"x":0.4288,"y":0.5922,"text":"Research focused on collaborative and competitive decision-making in multi-agent systems using reinforcement learning techniques."},{"id":"concept_t23","kind":"concept","n":24,"label":"Vision-Language Models","authors":"","year":null,"url":"","region":"t23","hemi":"wm","lobe":"occipital","color":"#4f799a","indegree":16,"alsoIn":[],"x":0.4115,"y":0.3172,"text":"Research focused on advancing vision-language models through pretraining, adaptation, and interpretability."},{"id":"concept_t24","kind":"concept","n":25,"label":"LLM Inference Optimization","authors":"","year":null,"url":"","region":"t24","hemi":"llm","lobe":"frontal","color":"#6784b0","indegree":11,"alsoIn":[],"x":0.8952,"y":0.5848,"text":"This cluster explores techniques to enhance the efficiency, throughput, and resource utilization of large language model inference systems."},{"id":"concept_t25","kind":"concept","n":26,"label":"Free Energy in Brain Theory","authors":"","year":null,"url":"","region":"t25","hemi":"llm","lobe":"parietal","color":"#6779b0","indegree":12,"alsoIn":[],"x":0.5859,"y":0.8215,"text":"This cluster explores the free energy principle as a unifying framework for understanding brain function, perception, and action through mathematical and theoretical lenses."},{"id":"concept_t26","kind":"concept","n":27,"label":"Model-Based Reinforcement Learning","authors":"","year":null,"url":"","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":25,"alsoIn":[],"x":0.2226,"y":0.3587,"text":"A cluster of papers exploring model-based approaches to reinforcement learning in robotics and embodied AI."},{"id":"concept_t27","kind":"concept","n":28,"label":"Efficient Reasoning in LLMs","authors":"","year":null,"url":"","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":23,"alsoIn":[],"x":0.8189,"y":0.8,"text":"This cluster explores methods to enhance reasoning efficiency in large language models through planning, reinforcement learning, and test-time compute optimization."},{"id":"concept_t28","kind":"concept","n":29,"label":"Memory and Cognitive Maps","authors":"","year":null,"url":"","region":"t28","hemi":"llm","lobe":"frontal","color":"#7667b0","indegree":15,"alsoIn":[],"x":0.5739,"y":0.4498,"text":"Exploring how the brain and algorithms represent and process abstract spatial and conceptual knowledge for planning and reasoning."},{"id":"concept_t29","kind":"concept","n":30,"label":"Video Generation & World Models","authors":"","year":null,"url":"","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":23,"alsoIn":[],"x":0.4229,"y":0.5568,"text":"Research exploring video generation techniques and their integration with world models for embodied AI and multi-modal understanding."},{"id":"concept_t30","kind":"concept","n":31,"label":"Multimodal Vision-Language Models","authors":"","year":null,"url":"","region":"t30","hemi":"llm","lobe":"temporal","color":"#8c67b0","indegree":12,"alsoIn":[],"x":0.7861,"y":0.2237,"text":"This cluster explores advancements and evaluations of multimodal models that integrate vision and language for embodied AI and reasoning tasks."},{"id":"concept_t31","kind":"concept","n":32,"label":"Retrieval-Augmented LLMs","authors":"","year":null,"url":"","region":"t31","hemi":"llm","lobe":"occipital","color":"#9767b0","indegree":9,"alsoIn":[],"x":0.8641,"y":0.6378,"text":"Research exploring the integration of retrieval mechanisms with large language models to enhance generation quality and task performance."},{"id":"concept_t32","kind":"concept","n":33,"label":"Diffusion Model Advancements","authors":"","year":null,"url":"","region":"t32","hemi":"wm","lobe":"frontal","color":"#8b4f9a","indegree":20,"alsoIn":[],"x":0.3917,"y":0.2142,"text":"This cluster explores improvements and applications of diffusion models in image and video generation, including training methods, efficiency, and integration with reinforcement learning."},{"id":"concept_t33","kind":"concept","n":34,"label":"LLM Evaluation Benchmarks","authors":"","year":null,"url":"","region":"t33","hemi":"llm","lobe":"parietal","color":"#ad67b0","indegree":12,"alsoIn":[],"x":0.8448,"y":0.6006,"text":"This cluster explores benchmarking frameworks and datasets to assess the capabilities and fairness of large language models and LLM-as-a-Judge systems."},{"id":"concept_t34","kind":"concept","n":35,"label":"Diffusion Weather Forecasting","authors":"","year":null,"url":"","region":"t34","hemi":"wm","lobe":"temporal","color":"#9a4f93","indegree":9,"alsoIn":[],"x":0.342,"y":0.1901,"text":"These papers explore using diffusion models for probabilistic and ensemble weather forecasting at global scales."},{"id":"concept_t35","kind":"concept","n":36,"label":"LLM Agent Evaluation","authors":"","year":null,"url":"","region":"t35","hemi":"llm","lobe":"occipital","color":"#b0679e","indegree":11,"alsoIn":[],"x":0.6494,"y":0.9284,"text":"This cluster explores methods and benchmarks for evaluating the capabilities and safety of large language model agents."},{"id":"concept_t36","kind":"concept","n":37,"label":"Cascading Dynamics in Networks","authors":"","year":null,"url":"","region":"t36","hemi":"llm","lobe":"frontal","color":"#b06793","indegree":8,"alsoIn":[],"x":0.6419,"y":0.6378,"text":"This cluster explores cascading phenomena across complex networks and turbulence, linking physical reasoning with network science and statistical physics."},{"id":"concept_t37","kind":"concept","n":38,"label":"Quantization Techniques for LLMs","authors":"","year":null,"url":"","region":"t37","hemi":"llm","lobe":"parietal","color":"#b06788","indegree":12,"alsoIn":[],"x":0.8731,"y":0.3767,"text":"Research focused on efficient quantization methods to reduce the size and improve the performance of large language models."},{"id":"concept_t38","kind":"concept","n":39,"label":"World Models for Physical AI","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":25,"alsoIn":[],"x":0.3988,"y":0.7802,"text":"Research focused on developing world models to enable spatial intelligence and physical reasoning in embodied AI systems."},{"id":"concept_t39","kind":"concept","n":40,"label":"Agentic Wireless Intelligence","authors":"","year":null,"url":"","region":"t39","hemi":"wm","lobe":"occipital","color":"#9a4f5b","indegree":8,"alsoIn":[],"x":0.2865,"y":0.4519,"text":"This cluster explores agentic and composable intelligence for next-generation wireless systems through world models, reinforcement learning, and distributed inference."},{"id":"arxiv:1708.00133","kind":"paper","n":41,"label":"Grounding Language for Transfer in Deep Reinforcement Learning","authors":"Narasimhan, Barzilay & Jaakkola","year":2018,"url":"https://arxiv.org/abs/1708.00133","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":3,"alsoIn":[],"x":0.1978,"y":0.2892,"text":"The paper studies how natural-language descriptions of environment entities can drive policy transfer in deep reinforcement learning across different game domains. The authors propose TEXT-VIN, a model-based agent combining a factorized state representation (entity-ID embedding + text-derived vector) with a differentiable Value Iteration Network and a model-free DQN component, trained end-to-end on source tasks and fine-tuned on target tasks. By grounding text to environment dynamics (transitions/rewards), the agent bootstraps learning in unseen domains without explicit inter-task entity mappings. Across GVGAI games it consistently beats baselines including Actor-Mimic, with up to 14% absolute higher average reward and 11.5% higher initial reward."},{"id":"arxiv:2603.00479","kind":"paper","n":42,"label":"U-VLM: Hierarchical Vision Language Modeling for Report Generation","authors":"","year":null,"url":"https://arxiv.org/abs/2603.00479","region":"t13","hemi":"llm","lobe":"parietal","color":"#6bb067","indegree":2,"alsoIn":[],"x":0.6346,"y":0.125,"text":"Automated 3D radiology report generation struggles because existing vision-language models don't use segmentation-pretrained encoders and inject visual features only at the language model's input layer, losing multi-scale detail. The paper proposes U-VLM, which trains a shared U-Net encoder through progressive stages (segmentation \u2192 classification \u2192 report generation, each usable with different datasets) and routes hierarchical U-Net encoder features to corresponding language-model layers via skip-connection-style multi-layer visual injection. Using only a 0.1B decoder trained from scratch, U-VLM reaches state-of-the-art results on CT-RATE (F1 0.414 vs BTB3D 0.258, BLEU-mean 0.349 vs 0.305) and AbdomenAtlas 3.0 lesion detection (F1 0.624 vs 0.518). Ablations show progressive pretraining mainly lifts F1 while multi-layer injection mainly lifts BLEU-mean, and that strong vision-encoder pretraining beats a 7B+ pretrained LLM decoder."},{"id":"arxiv:2305.10601","kind":"paper","n":43,"label":"Tree of Thoughts: Deliberate Problem Solving with Large Language Models","authors":"Yao et al.","year":2023,"url":"https://arxiv.org/abs/2305.10601","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":2,"alsoIn":[],"x":0.7988,"y":0.7947,"text":"Autoregressive LLMs make token-level, left-to-right decisions and struggle on tasks needing exploration, lookahead, or backtracking. The paper introduces Tree of Thoughts (ToT), an inference framework that generalizes Chain-of-Thought by maintaining a tree of intermediate 'thoughts', using the LM itself to generate and self-evaluate candidate states, and combining this with BFS/DFS search to deliberately explore, look ahead, and backtrack. Across three new tasks (Game of 24, Creative Writing, Mini Crosswords) ToT substantially outperforms IO and CoT prompting with GPT-4, e.g. raising Game of 24 success from 4% (CoT) to 74%. No extra training is required \u2014 it works with an off-the-shelf pre-trained LM."},{"id":"arxiv:2502.12130","kind":"paper","n":44,"label":"Scaling Autonomous Agents via Automatic Reward Modeling And Planning","authors":"Zhenfang Chen, Delin Chen, Rui Sun, Wenjun Liu, Chuang Gan","year":2025,"url":"https://arxiv.org/abs/2502.12130","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":4,"alsoIn":[],"x":0.6555,"y":0.8579,"text":"LLM agents struggle with multi-step decision-making tasks, and collecting trajectory data or fine-tuning closed-source policy LLMs is costly. ARMAP automatically learns a reward model from environment exploration without human annotations: an LLM agent navigates randomly to generate trajectories, another LLM synthesizes refined task intents plus positive/negative trajectory pairs, and a small VILA-3B vision-language model is trained as a reward model to score trajectories. This reward model is combined with planning algorithms (Best-of-N, Reflexion, MCTS) at inference time to guide frozen LLM agents. Across Webshop, ScienceWorld, Game of 24, ALFWorld and AgentClinic it consistently beats Sampling/Greedy baselines and even outperforms a 70B LLM used directly as reward model."},{"id":"arxiv:2510.26802","kind":"paper","n":45,"label":"Are Video Models Ready as Zero-Shot Reasoners? An Empirical Study with the MME-CoF Benchmark","authors":"Guo, Li et al.","year":2025,"url":"https://arxiv.org/abs/2510.26802","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":2,"alsoIn":[],"x":0.3745,"y":0.5791,"text":"The paper conducts the first systematic empirical study of whether modern video generation models can serve as zero-shot visual reasoners via 'Chain-of-Frame' (CoF) reasoning, focusing primarily on Veo-3. The authors curate MME-CoF, a compact benchmark of 59 cases spanning 12 reasoning dimensions (spatial, geometric, physical, temporal, embodied, etc.), and evaluate six leading models (Veo-3 variants, Sora-2 variants, Kling-v1, Seedance-1.0-pro) by generating six videos per prompt and grading them with Gemini-2.5-Pro on a 0\u20134 scale across five metrics plus a Good/Moderate/Bad rating and success rate. Findings show models handle short-horizon spatial coherence, fine-grained grounding, and locally consistent dynamics but fail at long-horizon causal reasoning, strict geometric constraints, and abstract logic. The conclusion is that current video models are not yet reliable standalone zero-shot reasoners, though they show promise as complementary visual engines alongside dedicated reasoning models."},{"id":"arxiv:2510.21219","kind":"paper","n":46,"label":"World Models Should Prioritize the Unification of Physical and Social Dynamics","authors":"","year":null,"url":"https://arxiv.org/abs/2510.21219","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":5,"alsoIn":[],"x":0.3663,"y":0.8883,"text":"This NeurIPS 2025 position paper argues that AI world models currently develop physical-dynamics prediction (e.g., MBRL, video generation, 3D models) and social-dynamics prediction (e.g., Theory of Mind, multi-agent RL, LLM social agents) in separate silos, leaving them unable to model the bidirectional interplay between physical environments and social behavior. The authors survey both dimensions through a dual physical-social lens, diagnose an 'integration gap' as a fundamental barrier rather than a missing feature, and propose the ACE Principles (Abstraction of social complexity, Contingent Causality, Entangled Emergence) to guide unification. They formalize a unified physical-social world model (WMP-S) as a joint state-transition problem and outline a three-tier hierarchical evaluation protocol plus a research roadmap. The central claim is that systematic, bidirectional unification of physical and social predictive capabilities is the next crucial frontier for world models."},{"id":"arxiv:2312.03815","kind":"paper","n":47,"label":"LLM as OS, Agents as Apps: Envisioning AIOS, Agents and the AIOS-Agent Ecosystem","authors":"Ge, Chen, Zhang et al.","year":2023,"url":"https://arxiv.org/abs/2312.03815","region":"t35","hemi":"llm","lobe":"occipital","color":"#b0679e","indegree":1,"alsoIn":[],"x":0.5861,"y":0.9477,"text":"This is a vision/position paper proposing the AIOS-Agent ecosystem, where a Large Language Model serves as an (Artificial) Intelligent Operating System (AIOS/LLMOS) and LLM-based AI agents act as applications, replacing the traditional OS-APP paradigm. It formalizes a conceptual framework that maps LLMOS components to classic OS elements (LLM\u2194kernel, context window\u2194memory, external storage\u2194file system, tools\u2194devices/libraries, user prompts\u2194user interface) and argues natural language becomes the programming interface. The paper surveys single-agent, multi-agent, and human-agent applications across many domains, then derives OS-inspired future research directions in resource/memory management, communication protocols (DSLs, 'LLM as Interpreter'), and security. No empirical experiments are run; the contribution is the framework, taxonomy, and roadmap."},{"id":"arxiv:2402.01695","kind":"paper","n":48,"label":"Language-Guided World Models: A Model-Based Approach to AI Control","authors":"Zhang, Nguyen, Tuyls, Lin & Narasimhan","year":2024,"url":"https://arxiv.org/abs/2402.01695","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":5,"alsoIn":[],"x":0.7123,"y":0.7988,"text":"The paper introduces Language-Guided World Models (LWMs), probabilistic world models that simulate environment dynamics by reading natural-language manuals, letting humans control model-based agents through verbal communication. The authors build MESSENGER-WM, a benchmark on the MESSENGER grid-world game with three settings (NewCombo, NewAttr, NewAll) testing increasing compositional generalization. They show a standard Transformer world model barely beats a no-text baseline, then propose EMMA-LWM, which fuses the Transformer with EMMA two-step attention to ground entity identities to attribute descriptions and substantially outperforms baselines, approaching an oracle-parsing skyline. A plan-discussion application shows the model lets an agent present and revise plans from language feedback, achieving 3-4x higher reward than an observational world model."},{"id":"arxiv:2311.05348","kind":"paper","n":49,"label":"u-LLaVA: Unifying Multi-Modal Tasks via Large Language Model","authors":"Jinjin Xu, Liwu Xu, Yuzhe Yang, Xiang Li, Yanchun Xie, et al.","year":2023,"url":"https://arxiv.org/abs/2311.05348","region":"t30","hemi":"llm","lobe":"temporal","color":"#8c67b0","indegree":2,"alsoIn":[],"x":0.8111,"y":0.2123,"text":"u-LLaVA is a unifying multi-modal LLM framework that jointly handles global (image/video), regional (bounding-box), and pixel-level (mask) understanding within a single instruction-tuning stage. It uses an efficient two-stage approach: Stage I aligns CLIP ViT-L/14 image and spatio-temporal video features to a Vicuna LLM, and Stage II adds task-specific projectors and decoders connecting the LLM to a SAM mask decoder (pixel) and an MLP location decoder (region). The authors also release ullava-277K, a 277K-sample mask-based multi-task dataset (including a reconstructed Salient-15K). u-LLaVA-7B achieves SOTA among 7B MLLMs on RES, salient segmentation, and REC benchmarks, even surpassing some specialist expert models while using ~1/10 of their training data."},{"id":"arxiv:2401.12945","kind":"paper","n":50,"label":"Lumiere: A Space-Time Diffusion Model for Video Generation","authors":"Bar-Tal et al. (Google)","year":2024,"url":"https://arxiv.org/abs/2401.12945","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":0,"alsoIn":[],"x":0.4311,"y":0.3229,"text":"Lumiere addresses the difficulty of generating globally coherent motion in text-to-video diffusion, which prior cascaded models (base keyframes + temporal super-resolution) handle poorly due to temporal aliasing and fixed small context windows. It introduces a Space-Time U-Net (STUNet) that generates the entire video duration in a single pass by downsampling/upsampling in both space and time, built on a frozen pretrained text-to-image diffusion model, and extends MultiDiffusion along the temporal axis for memory-feasible spatial super-resolution. The model generates 80 frames at 16fps (5 seconds) and achieves competitive UCF101 zero-shot FVD/IS while being preferred over baselines in user studies, and supports image-to-video, inpainting, stylization, and cinemagraphs."},{"id":"arxiv:2504.00983","kind":"paper","n":51,"label":"WorldScore: A Unified Evaluation Benchmark for World Generation","authors":"Duan et al.","year":2025,"url":"https://arxiv.org/abs/2504.00983","region":"t15","hemi":"wm","lobe":"occipital","color":"#4f9a62","indegree":0,"alsoIn":[],"x":0.4181,"y":0.7312,"text":"Existing video benchmarks (e.g., VBench) only assess single-scene generation and lack camera/reference-image specifications, making them incompatible with 3D/4D scene generation methods and unable to measure multi-scene world generation. WorldScore decomposes world generation into a sequence of next-scene tasks defined by a triplet (current scene, next scene, layout) with explicit camera-trajectory layouts and a standardized video output format, enabling unified evaluation across 3D, 4D, T2V, and I2V models on three aspects\u2014controllability, quality, and dynamics\u2014via 10 metrics. The benchmark provides 3,000 curated test examples (2,000 static, 1,000 dynamic; photorealistic and stylized, indoor/outdoor). Evaluating 20 models reveals that 3D models excel at static worlds (WonderWorld 72.69) while video models lag badly in camera controllability and long-sequence/outdoor scenes, and the best open-source video model (CogVideoX-I2V) beats both closed-source models."},{"id":"arxiv:1910.10683","kind":"paper","n":52,"label":"Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer","authors":"Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael ","year":2020,"url":"https://arxiv.org/abs/1910.10683","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":3,"alsoIn":[],"x":0.8215,"y":0.5575,"text":"The paper conducts a large-scale systematic study of transfer learning for NLP by casting every text task (translation, QA, classification, summarization, regression) into a unified text-to-text format, enabling the same model, loss, and decoding to be applied everywhere. It introduces the 750GB Colossal Clean Crawled Corpus (C4) and uses a controlled 'coordinate ascent' methodology to compare architectures, pre-training objectives, datasets, transfer/multi-task strategies, and scaling against a ~220M-parameter baseline. Combining the best insights (encoder-decoder, span-corruption denoising, longer training, multi-task pre-training) with scale up to 11 billion parameters, the resulting T5 model achieves state-of-the-art on 18 of 24 benchmarks, including a GLUE average of 90.3 and a SuperGLUE average of 88.9 that nearly matches human performance."},{"id":"arxiv:2512.04797","kind":"paper","n":53,"label":"SIMA 2: A Generalist Embodied Agent for Virtual Worlds","authors":"SIMA Team (Google DeepMind)","year":2025,"url":"https://arxiv.org/abs/2512.04797","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":1,"alsoIn":[],"x":0.2574,"y":0.7237,"text":"SIMA 2 is a generalist embodied agent built on a Gemini Flash-Lite foundation model that perceives 3D virtual worlds via RGB pixels and acts through a keyboard-and-mouse interface, while also reasoning, conversing, and handling multi-modal (text+image/sketch) instructions. It is trained via supervised finetuning on a mixture of human gameplay trajectories and synthetic 'bridge' data (reasoning/dialogue annotated by Gemini), then refined with online RL from verifiable rewards. SIMA 2 roughly doubles SIMA 1's task success and approaches human performance across a portfolio of commercial and research games, generalizes to entirely held-out environments (ASKA, Minecraft/MineDojo, The Gunk, and Genie 3 photorealistic worlds), and retains the base model's reasoning ability. It further demonstrates open-ended self-improvement: using Gemini as both task setter and reward model, the agent autonomously learns new skills from self-generated experience in unseen environments."},{"id":"psychclassics.yorku.ca:d7323f879ff8","kind":"paper","n":54,"label":"Cognitive maps in rats and men","authors":"Tolman, E. C.","year":1948,"url":"https://psychclassics.yorku.ca/Tolman/Maps/maps.htm","region":"t5","hemi":"llm","lobe":"parietal","color":"#b09e67","indegree":0,"alsoIn":[],"x":0.6046,"y":0.5118,"text":"Tolman synthesizes a body of rat maze experiments to argue against the dominant stimulus-response (telephone-switchboard) theory of learning and in favor of 'field theory': that rats build an internal, map-like representation of their environment in the brain. He marshals five experiment types \u2014 latent learning, vicarious trial and error (VTE), searching for the stimulus, hypotheses, and spatial orientation \u2014 to show learning is active, selective, and spatial rather than passive chaining of reflexes. He further distinguishes narrow 'strip maps' from broad 'comprehensive maps,' showing rats can take novel shortcuts toward a goal's true direction. He closes by extrapolating that overly strong motivation or frustration narrows cognitive maps, offering an account of human maladjustments like regression, fixation, and displaced aggression."},{"id":"arxiv:2311.07222","kind":"paper","n":55,"label":"Neural General Circulation Models for Weather and Climate (NeuralGCM)","authors":"Kochkov et al.","year":2023,"url":"https://arxiv.org/abs/2311.07222","region":"t34","hemi":"wm","lobe":"temporal","color":"#9a4f93","indegree":3,"alsoIn":[],"x":0.2548,"y":0.1381,"text":"Weather and climate prediction has relied on physics-based general circulation models (GCMs) that are computationally expensive and need hand-tuned parameterizations, while recent pure-ML forecasters lack calibrated ensembles and long-term stability. The paper presents NeuralGCM, the first fully-differentiable hybrid GCM that couples a JAX-based differentiable dynamical core (solving the hydrostatic primitive equations) with a neural-network learned-physics module, trained end-to-end ('online') on ERA5 trajectories of up to 5 days. NeuralGCM matches the best ML and physics-based models for 1-10 day deterministic forecasts and beats ECMWF-ENS on ensemble CRPS for 1-15 day forecasts, while remaining stable for decadal climate runs with prescribed SST and reproducing emergent phenomena like realistic tropical-cyclone tracks. It achieves these results at 8-40x coarser resolution, yielding 3-5 orders of magnitude in compute savings over conventional GCMs."},{"id":"arxiv:2605.12090","kind":"paper","n":56,"label":"World Action Models: The Next Frontier in Embodied AI","authors":"Siyin Wang, et al.","year":2026,"url":"https://arxiv.org/abs/2605.12090","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":7,"alsoIn":[],"x":0.1919,"y":0.7951,"text":"This survey introduces and formally defines 'World Action Models' (WAMs)\u2014embodied foundation models that unify predictive world-state modeling with action generation, targeting a joint distribution p(o',a|o,l) over future states and actions rather than reactive observation-to-action mappings as in standard VLA models. It organizes the fragmented literature into a taxonomy of Cascaded WAMs (explicit pixel-space or implicit latent planning, then action decoding) and Joint WAMs (autoregressive or diffusion-based, further split into unified-stream and multi-stream architectures). It catalogs the four-pillar data ecosystem (robot teleoperation, portable UMI-style human demos, simulation, internet-scale egocentric video) and synthesizes evaluation protocols around visual fidelity, physical commonsense, and action plausibility. It concludes with open challenges including architectural coupling, multimodal state representation, data-mixture design, long-horizon planning, inference latency, joint evaluation, and safety."},{"id":"arxiv:2506.18897","kind":"paper","n":57,"label":"MinD: Unified Visual Imagination and Control via Hierarchical World Models","authors":"","year":null,"url":"https://arxiv.org/abs/2506.18897","region":"t17","hemi":"llm","lobe":"parietal","color":"#4f9a79","indegree":7,"alsoIn":[],"x":0.6495,"y":0.3307,"text":"VGMs are typically used only as frozen feature-extraction backbones for VLA policies, ignoring their generative future-prediction capability, and naive multi-step video diffusion is too slow for real-time control. MinD is a dual-system hierarchical world model that runs a slow low-frequency video generator (LoDiff-Visual) and a fast high-frequency diffusion action policy (HiDiff-Policy), conditioning actions on a single-step (single denoising timestep) latent rather than a fully denoised video, bridged by a DiffMatcher module trained with a diffusion-forcing alignment loss. It achieves 63% mean success on RLBench and ~69-72% on a real-world Franka at 11.3 FPS, and as a secondary finding identifies 74% of task failures in advance from generated video clips. This establishes generative VGMs as efficient, interpretable, risk-aware world models for manipulation."},{"id":"arxiv:2511.00062","kind":"paper","n":58,"label":"World Simulation with Video Foundation Models for Physical AI","authors":"NVIDIA (Arslan Ali et al.)","year":2025,"url":"https://arxiv.org/abs/2511.00062","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":7,"alsoIn":[],"x":0.4375,"y":0.9478,"text":"NVIDIA introduces Cosmos-Predict2.5, a flow-matching video world foundation model for Physical AI that unifies Text2World, Image2World, and Video2World in one network and replaces the T5 text encoder with the Cosmos-Reason1 VLM for richer grounding. Trained on 200M curated clips (filtered from ~6B, a 4% survival rate) and refined with SFT, model merging, and GRPO-style reinforcement learning, it is released at 2B and 14B scales and matches or beats much larger Wan2.2 models on PAI-Bench. The paper also presents Cosmos-Transfer2.5, a control-net variant that is 3.5\u00d7 smaller than Cosmos-Transfer1-7B yet delivers higher fidelity and far less error accumulation in long-horizon generation. The models are validated as synthetic-data generators and simulators across robot policy learning, autonomous-driving multi-view simulation, camera-controlled multi-view generation, VLA training, and action-conditioned world generation."},{"id":"arxiv:2601.15533","kind":"paper","n":59,"label":"From Generative Engines to Actionable Simulators: The Imperative of Physical Grounding in World Models","authors":"Zhikang Chen, Tingting Zhu","year":2026,"url":"https://arxiv.org/abs/2601.15533","region":"t17","hemi":"llm","lobe":"parietal","color":"#4f9a79","indegree":19,"alsoIn":[],"x":0.669,"y":0.5932,"text":"This survey/position paper argues that current world models suffer from 'visual conflation' \u2014 the mistaken assumption that high-fidelity video generation implies an understanding of physical and causal dynamics. It contends that visual realism is a necessary but insufficient proxy for world understanding, since photorealistic generators frequently violate invariant constraints, fail under intervention, and break down in safety-critical decisions. The authors reframe world models as 'actionable simulators' rather than visual engines, organizing recent work around four challenges (structured 4D interfaces, self-evolution, physical anchoring, and generalization through imagination) and calling for closed-loop, decision-oriented evaluation. They use medical decision-making as an epistemic stress test to show a world model's value lies in supporting counterfactual reasoning, intervention planning, and long-horizon foresight rather than realistic-looking rollouts."},{"id":"arxiv:2510.04999","kind":"paper","n":60,"label":"Bridging Text and Video Generation: A Survey","authors":"Nilay Kumar et al.","year":2025,"url":"https://arxiv.org/abs/2510.04999","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":5,"alsoIn":[],"x":0.4441,"y":0.3594,"text":"This survey reviews text-to-video (T2V) generation models from early GAN- and VAE-based approaches to modern diffusion-transformer (DiT) architectures, covering 17 representative models (MoCoGAN through Pyramidal Flow). It systematically catalogs training datasets (e.g., WebVid-10M, UCF-101, HowTo100M), training configurations (GPU types, batch sizes, learning rates, optimizers), and evaluation metrics (IS, FID, FVD, CLIP-SIM, KVD) with benchmark performance comparisons. The paper also discusses limitations of current metrics and highlights VBench as a promising holistic evaluation framework with 16 dimensions of assessment. Open challenges include dataset scarcity, computational cost, temporal coherence, and physics realism."},{"id":"arxiv:2304.12995","kind":"paper","n":61,"label":"AudioGPT: Understanding and Generating Speech, Music, Sound, and Talking Head","authors":"Rongjie Huang et al.","year":2023,"url":"https://arxiv.org/abs/2304.12995","region":"t11","hemi":"llm","lobe":"occipital","color":"#81b067","indegree":6,"alsoIn":[],"x":0.8739,"y":0.2674,"text":"Current LLMs like ChatGPT cannot process complex audio or conduct spoken conversations, and training multimodal audio LLMs from scratch is data- and compute-prohibitive. AudioGPT addresses this by treating ChatGPT as a general-purpose interface that orchestrates a library of existing audio foundation models (for speech, music, sound, and talking head) plus an ASR/TTS input-output interface to enable spoken dialogue, running through four stages: modality transformation, task analysis, model assignment, and response generation. The system supports 15+ audio tasks across understanding and generation, and the authors propose an evaluation protocol along three axes \u2014 consistency, capability, and robustness. The paper is largely a system description with qualitative demonstrations (e.g., a 12-round multi-turn dialogue) rather than quantitative benchmark results."},{"id":"arxiv:2602.22010","kind":"paper","n":62,"label":"World Guidance: World Modeling in Condition Space for Action Generation","authors":"","year":null,"url":"https://arxiv.org/abs/2602.22010","region":"t17","hemi":"llm","lobe":"parietal","color":"#4f9a79","indegree":7,"alsoIn":[],"x":0.6465,"y":0.3215,"text":"VLA models that predict future observations to aid action generation face a trade-off: rich task-agnostic future representations carry redundancy and cost, while compact latent action spaces are too coarse for fine-grained control. WoG (World Guidance) maps future observations into a compact 'condition space' by injecting them (via a Q-Former-based Future Encoder over frozen DINOv2/Wan VAE/SigLIP features) into the action inference pipeline during stage I, then in stage II freezes the encoder and trains the VLM to jointly predict these future conditions and actions, making it self-guided at inference. Across SIMPLER simulation and real-world UR5 tasks WoG outperforms conventional VLAs, latent action models, and video-prediction world models (e.g. 69.4% overall on Google Robot vs 60.5% for \u03c00-FAST), and it scales by learning from large-scale human manipulation videos and UMI data."},{"id":"arxiv:2403.15498","kind":"paper","n":63,"label":"Emergent World Models and Latent Variable Estimation in Chess-Playing Language Models","authors":"Karvonen","year":2024,"url":"https://arxiv.org/abs/2403.15498","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":4,"alsoIn":[],"x":0.6959,"y":0.7526,"text":"The paper investigates whether language models build internal world models by training 8-layer (25M) and 16-layer (50M) character-level GPT models on next-token prediction over 16 million real human chess games from Lichess. Using linear probes on the residual stream, the author recovers an accurate internal board-state representation and, because games are real rather than synthetic, also recovers a latent player-skill (Elo) estimate. Causal interventions validate these representations: adding a derived skill vector raises win rate against Stockfish by up to 2.6x, and board-state edits make the model play legally under hypothetically modified boards. This extends Li et al.'s OthelloGPT result from synthetic to non-synthetic data and shows latent variable estimation emerges from next-character prediction."},{"id":"arxiv:2510.05865","kind":"paper","n":64,"label":"The Safety Challenge of World Models for Embodied AI Agents: A Review","authors":"","year":null,"url":"https://arxiv.org/abs/2510.05865","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":2,"alsoIn":[],"x":0.3462,"y":0.8483,"text":"This review surveys World Models (WMs) for embodied AI in autonomous driving and robotics, organizing them by scene generation vs. control tasks and by generative framework (diffusion vs. autoregressive). Its central contribution is a safety lens: it defines a set of 'pathology criteria' (visual quality, temporal consistency, traffic adherence, physical conformity, condition consistency, grasp consistency) for diagnosing unsafe faults in WM predictions. It complements the literature review with an empirical study, running SoTA WMs (Panacea, Vista, MagicDriveDiT, Cosmos, Open-Sora Plan, This&That for scenes; MILE, CTG++, LCTGen, Octo, RT-1-X, RoboGen for control) and scoring their outputs. All scene generators scored below 3 on a 1\u20134 scale, and robot controllers showed substantial physical-conformity and grasp failures, exposing a clear safety gap in current WMs."},{"id":"proceedings.mlr.press:6beb908867f9","kind":"paper","n":65,"label":"Can Large Language Models Reason about Program Invariants?","authors":"Pei et al.","year":2023,"url":"https://proceedings.mlr.press/v202/pei23a.html","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":0,"alsoIn":[],"x":0.814,"y":0.7416,"text":"The paper studies whether large language models can predict program invariants statically, replacing the dynamic analysis used by existing tools (e.g., Daikon) that require traces from multiple program executions. The authors fine-tune models pretrained on source code to generate invariants directly from program text, and compare prompting/decoding strategies including a sequential 'scratchpad' approach. They find that predicting invariants sequentially through the program with a scratchpad yields the best performance, producing static invariants of quality comparable to a dynamic analysis tool given five program traces."},{"id":"arxiv:2310.02207","kind":"paper","n":66,"label":"Language Models Represent Space and Time","authors":"Gurnee & Tegmark","year":2023,"url":"https://arxiv.org/abs/2310.02207","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":2,"alsoIn":[],"x":0.7258,"y":0.5272,"text":"The paper investigates whether LLMs learn coherent world models versus surface statistics by probing internal activations for spatial and temporal information. The authors build six datasets (world/USA/NYC places; historical figures, art/entertainment, news headlines) and train linear ridge-regression probes on Llama-2 (7B/13B/70B) and Pythia activations to predict latitude/longitude or timestamps. They find space and time are linearly decodable, improving with model scale and depth (plateauing around the model halfway point), robust to prompting, and unified across entity types. They further identify individual 'space neurons' and 'time neurons' that encode coordinates and show causal interventions on a time neuron change next-token predictions, suggesting LLMs possess basic ingredients of a world model."},{"id":"arxiv:2603.27918","kind":"paper","n":67,"label":"Adversarial Attacks on Multimodal Large Language Models: A Comprehensive Survey","authors":"","year":null,"url":"https://arxiv.org/abs/2603.27918","region":"t8","hemi":"llm","lobe":"frontal","color":"#a2b067","indegree":5,"alsoIn":[],"x":0.7766,"y":0.4427,"text":"This survey systematically analyzes adversarial attacks against multimodal large language models (MLLMs), moving beyond cataloging techniques to explain why models are susceptible. It introduces a goal-driven taxonomy that organizes attacks by adversarial objective into four families\u2014integrity, safety/jailbreak, control/injection, and poisoning/backdoor\u2014cross-cut by an attacker-knowledge axis (white/gray/black-box). It then presents a vulnerability-centric analysis mapping these attack families to shared architectural and representational weaknesses (cross-modal misalignment, embedding-space fragility, modality-specific processing, instruction-following ambiguity, training-data integrity), and surveys layered defenses. The synthesis covers 88 works, 65 with full empirical characterization consolidated in a summary table."},{"id":"arxiv:2602.15400","kind":"paper","n":68,"label":"One Agent to Guide Them All: Empowering MLLMs for Vision-and-Language Navigation via Explicit World Representation","authors":"Zerui Li, Hongpei Zheng, Fangguo Zhao, Aidan Chan, et al.","year":2026,"url":"https://arxiv.org/abs/2602.15400","region":"t12","hemi":"llm","lobe":"frontal","color":"#5e9a4f","indegree":2,"alsoIn":[],"x":0.6346,"y":0.2075,"text":"The paper addresses Vision-and-Language Navigation in Continuous Environments (VLN-CE), where tightly coupled MLLM designs suffer from error propagation between spatial and semantic reasoning. The authors propose GTA, a decoupled zero-shot framework that separates low-level spatial state estimation (via a TSDF volumetric map and topological graph) from high-level semantic planning (via a frozen MLLM performing counterfactual reasoning over an interactive metric world representation). GTA achieves new zero-shot state-of-the-art with 48.8% SR on R2R-CE and 42.2% SR on RxR-CE, and demonstrates zero-shot sim-to-real transfer on a TurtleBot 4 wheeled robot and a custom aerial drone."},{"id":"arxiv:2403.06845","kind":"paper","n":69,"label":"DriveDreamer-2: LLM-Enhanced World Models for Diverse Driving Video Generation","authors":"Zhao et al.","year":2024,"url":"https://arxiv.org/abs/2403.06845","region":"t15","hemi":"wm","lobe":"occipital","color":"#4f9a62","indegree":1,"alsoIn":[],"x":0.4132,"y":0.7582,"text":"DriveDreamer-2 extends the DriveDreamer world-model framework with a fine-tuned LLM (GPT-3.5) that converts free-text user queries into agent trajectories, a diffusion-based HDMap generator that produces traffic-rule-compliant road structures conditioned on those trajectories, and a Unified Multi-View Video Model (UniMVM) that unifies intra- and cross-view consistency for multi-view driving video generation. It is presented as the first world model to generate customized/uncommon driving videos (e.g., vehicles cutting in) from text prompts alone. On the nuScenes validation set it reaches FID 11.2 and FVD 55.7, relative improvements of ~30% and ~50% over prior best methods. The generated videos also augment training data, improving downstream StreamPETR 3D detection and tracking metrics."},{"id":"arxiv:2507.00917","kind":"paper","n":70,"label":"A Survey: Learning Embodied Intelligence from Physical Simulators and World Models","authors":"(survey authors)","year":2025,"url":"https://arxiv.org/abs/2507.00917","region":"t7","hemi":"wm","lobe":"occipital","color":"#adb067","indegree":9,"alsoIn":[],"x":0.2678,"y":0.881,"text":"A comprehensive survey (covering 2018\u20132025) of how embodied intelligence is learned through the integration of two complementary technologies: physical simulators (explicit external environments for training/testing) and world models (internal generative representations for prediction and planning). It proposes a five-level capability grading scheme for intelligent robots (IR-L0 to IR-L4), reviews robot locomotion/manipulation/human-robot interaction, gives a comparative analysis of mainstream simulators (Webots, Gazebo, MuJoCo, PyBullet, CoppeliaSim, Isaac Gym/Sim/Lab, SAPIEN, Genesis, Newton), and taxonomizes world-model architectures and their roles as neural simulators, dynamic models, and reward models. It then surveys world models specialized for autonomous driving and articulated robots, ending with open challenges (causal reasoning, evaluation, memory, interpretability)."},{"id":"arxiv:2402.15391","kind":"paper","n":71,"label":"Genie: Generative Interactive Environments","authors":"Bruce et al.","year":2024,"url":"https://arxiv.org/abs/2402.15391","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":5,"alsoIn":[],"x":0.3326,"y":0.639,"text":"Genie is the first generative interactive environment, an 11B-parameter foundation world model trained fully unsupervised from over 200,000 hours (filtered to 30k hours / 6.8M 16s clips) of unlabelled Internet 2D-platformer gameplay videos with no action or text labels. It combines a spatiotemporal (ST-transformer) video tokenizer, an unsupervised latent action model (VQ-VAE codebook of 8 actions), and an autoregressive MaskGIT dynamics model, enabling frame-by-frame action-controllable generation of playable worlds from text, sketch, or photo prompts. The architecture scales gracefully from 40M to 2.7B (final 10.1B) dynamics parameters, generalizes to a robotics dataset, and its learned latent actions transfer to imitate behaviors in unseen RL environments (CoinRun) with as few as 200 ground-truth action labels."},{"id":"arxiv:2102.01192","kind":"paper","n":72,"label":"Generative Spoken Language Modeling from Raw Audio (GSLM)","authors":"Kushal Lakhotia et al.","year":2021,"url":"https://arxiv.org/abs/2102.01192","region":"t11","hemi":"llm","lobe":"occipital","color":"#81b067","indegree":3,"alsoIn":[],"x":0.8515,"y":0.3164,"text":"The paper introduces Generative Spoken Language Modeling (GSLM), the task of learning a language's acoustic and linguistic structure from raw audio with no text or labels, plus a suite of automatic and human evaluation metrics. The baseline pipeline chains a self-supervised discrete speech encoder (speech-to-unit), a Transformer unit language model trained on the pseudo-text, and a Tacotron-2 + WaveGlow unit-to-speech synthesizer. Factorially crossing three encoders (CPC, wav2vec 2.0, HuBERT) with three codebook sizes (50/100/200), the authors show it is possible to train an LM from audio-derived units and generate intelligible, locally coherent speech, with some configurations approaching text-based toplines. They also validate that proposed ASR-based generation metrics correlate strongly with human judgments and can be proxied by cheaper zero-shot encoding metrics."},{"id":"arxiv:1705.09156","kind":"paper","n":73,"label":"The free energy principle for action and perception: A mathematical review","authors":"","year":null,"url":"https://arxiv.org/abs/1705.09156","region":"t25","hemi":"llm","lobe":"parietal","color":"#6779b0","indegree":2,"alsoIn":[],"x":0.5732,"y":0.8386,"text":"The free energy principle (FEP) proposes that biological organisms maintain themselves by minimizing a tractable upper bound \u2014 informational free energy (IFE) \u2014 on the surprisal (negative log-probability) of their sensory data, unifying perception, action, and learning. This paper supplies a complete, self-contained mathematical derivation of the predictive-coding implementation of the FEP: from the recognition (R) and generative (G) densities and the Laplace (Gaussian) approximation, through dynamic generative models in generalized coordinates and hierarchical models, up to a combined 'full construct'. It grounds the formalism in a simple agent-based 'Bayesian thermostat', showing how minimizing IFE unifies perceptual inference (updating brain states) and active inference (acting on the world). Finally it explicitly catalogs the non-obvious assumptions and approximations (Gaussianity, tightly-peaked densities, local linearity, gradient descent, hard-wired inverse models) that the framework depends on."},{"id":"arxiv:1706.03741","kind":"paper","n":74,"label":"Deep Reinforcement Learning from Human Preferences","authors":"Christiano et al.","year":2017,"url":"https://arxiv.org/abs/1706.03741","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":6,"alsoIn":[],"x":0.2315,"y":0.206,"text":"Addresses the problem of communicating complex, hard-to-specify goals to deep RL agents without access to a hand-engineered reward function. The method learns a reward function from non-expert human preferences over pairs of short (1-2 second) trajectory-segment video clips, fitting it via a Bradley-Terry/Elo-style cross-entropy model while simultaneously optimizing a policy (A2C for Atari, TRPO for MuJoCo) on the predicted reward, with asynchronous online query selection by ensemble disagreement. It solves most Atari games and MuJoCo locomotion tasks using feedback on less than 1% of agent interactions, and trains novel behaviors (e.g., Hopper backflips) from about an hour of human time. This reduces interaction complexity by roughly 3 orders of magnitude, making human oversight economically practical for state-of-the-art RL."},{"id":"arxiv:2204.03458","kind":"paper","n":75,"label":"Video Diffusion Models","authors":"Jonathan Ho et al.","year":2022,"url":"https://arxiv.org/abs/2204.03458","region":"t32","hemi":"wm","lobe":"frontal","color":"#8b4f9a","indegree":14,"alsoIn":[],"x":0.3922,"y":0.342,"text":"The paper introduces Video Diffusion Models (VDM), extending the standard Gaussian image diffusion model to video using a space-time factorized 3D U-Net (space-only 3D convolutions plus interleaved spatial and temporal attention blocks with relative position embeddings). It enables joint training on image and video objectives (by appending masked independent image frames), which reduces minibatch gradient variance, and introduces a 'reconstruction-guided' conditional sampling method to autoregressively extend videos and perform spatial/temporal super-resolution. The model achieves state-of-the-art sample-quality scores on unconditional UCF101 and on the BAIR and Kinetics-600 video-prediction benchmarks, and presents the first results on a large (10M caption) text-conditioned video generation task."},{"id":"arxiv:2507.10535","kind":"paper","n":76,"label":"CodeJudgeBench: Benchmarking LLM-as-a-Judge for Coding Tasks","authors":"Hongchao Jiang et al.","year":2025,"url":"https://arxiv.org/abs/2507.10535","region":"t33","hemi":"llm","lobe":"parietal","color":"#ad67b0","indegree":4,"alsoIn":[],"x":0.8933,"y":0.7048,"text":"Despite the growing use of LLM-as-a-Judge, no dedicated benchmark existed for evaluating judges on coding tasks, so the authors introduce CodeJudgeBench, a 5,352-pair execution-free benchmark spanning code generation, code repair, and unit test generation, with challenging chosen/rejected pairs sourced from strong models (Gemini-2.5-Pro, Claude-3.7-Sonnet) on LiveCodeBench-v6 problems. Benchmarking 26 LLM judges, they find recent thinking models substantially outperform non-thinking and specially-tuned judge models\u2014e.g., Qwen3-8B beats 70B judge-tuned models. However, all judges show high randomness: pairwise position swaps shift accuracy by up to 14%, and performance varies by which model produced the responses. They also find pairwise comparison beats point-wise scalar scoring and that feeding the full unprocessed response (with comments/reasoning) yields better judging than code-only."},{"id":"arxiv:2005.14165","kind":"paper","n":77,"label":"Language Models are Few-Shot Learners","authors":"Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla","year":2020,"url":"https://arxiv.org/abs/2005.14165","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":4,"alsoIn":[],"x":0.814,"y":0.6176,"text":"The paper investigates whether scaling autoregressive language models dramatically improves task-agnostic few-shot performance without any gradient updates or fine-tuning. The authors train GPT-3, a 175-billion-parameter transformer (10x larger than any prior non-sparse LM) on ~300B tokens, and evaluate it on two dozen+ NLP benchmarks plus novel synthetic tasks under zero-, one-, and few-shot in-context learning, where tasks and demonstrations are specified purely as text in the 2048-token context window. GPT-3 reaches or surpasses fine-tuned SOTA on several tasks (e.g., LAMBADA, TriviaQA, PhysicalQA) while struggling on others (ANLI, WiC, RACE), and the few-shot gap over zero-shot grows with model size. It also generates news articles humans can barely distinguish from real ones (~52% detection accuracy) and the authors analyze data contamination, bias, and broader societal impacts."},{"id":"arxiv:2512.00597","kind":"paper","n":78,"label":"Scaling Down to Scale Up: Towards Operationally-Efficient and Deployable Clinical Models via Cross-Modal Low-Rank Adaptation for Medical Vision-Language Models","authors":"","year":null,"url":"https://arxiv.org/abs/2512.00597","region":"t13","hemi":"llm","lobe":"parietal","color":"#6bb067","indegree":5,"alsoIn":[],"x":0.7438,"y":0.2299,"text":"Adapting large CT vision-language foundation models to clinical tasks normally requires full fine-tuning of hundreds of millions of parameters, which is costly, storage-heavy, and risks catastrophic forgetting. The authors introduce MedCT-VLM, which inserts Low-Rank Adaptation (LoRA) modules into the attention layers of both the vision (CTViT) and text (BiomedVLP-CXR-BERT) encoders of CT-CLIP, training only 1.67M parameters (0.38% of the 440M total) while keeping the base frozen. Evaluated on zero-shot classification of 18 thoracic pathologies from CT-RATE, LoRA fine-tuning raises mean AUROC from 61.3% to 68.9% (+7.6 pp), accuracy from 67.2% to 73.6%, and macro-F1 from 32.1% to 36.9%. The method reduces checkpoint size by 74x and enables multi-task deployment via adapter swapping."},{"id":"arxiv:2210.03629","kind":"paper","n":79,"label":"ReAct: Synergizing Reasoning and Acting in Language Models","authors":"Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, Yu","year":2023,"url":"https://arxiv.org/abs/2210.03629","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":1,"alsoIn":[],"x":0.7378,"y":0.8894,"text":"The paper addresses the gap between reasoning (chain-of-thought) and acting (action-plan generation) in LLMs, which had been studied separately. It introduces ReAct, a prompting paradigm that interleaves free-form verbal reasoning traces with task-specific actions, letting the model build/track/update plans (reason to act) while gathering external information from environments like a Wikipedia API (act to reason). Using few-shot prompting on a frozen PaLM-540B, ReAct reduces hallucination on QA/fact-verification and beats imitation/RL baselines on interactive decision-making by absolute success-rate margins of 34% (ALFWorld) and 10% (WebShop) with only 1-2 in-context examples. Combining ReAct with CoT self-consistency yields the best overall reasoning results, and finetuning shows ReAct generalizes better than reasoning- or acting-only baselines."},{"id":"arxiv:2202.09467","kind":"paper","n":80,"label":"Predictive Coding: Towards a Future of Deep Learning beyond Backpropagation?","authors":"Beren Millidge, Tommaso Salvatori, Yuhang Song, Rafal Bogacz, Thomas Lukasiewicz","year":2022,"url":"https://arxiv.org/abs/2202.09467","region":"t20","hemi":"llm","lobe":"frontal","color":"#67b0b0","indegree":6,"alsoIn":[],"x":0.6271,"y":0.6508,"text":"This survey reviews predictive coding (PC), a neuroscience-derived error-driven learning algorithm with purely local weight updates, as a potential alternative to backpropagation (BP) for training deep neural networks. It synthesizes recent theoretical results showing PC can approximate or (via the Z-IL variant) exactly replicate BP's parameter updates on MLPs and arbitrary computational graphs, and surveys empirical results where predictive coding networks (PCNs) match BP on classification and generation benchmarks. It emphasizes PCNs' greater flexibility \u2014 a single network can act as classifier, generator, and associative memory and be defined on arbitrary graph topologies \u2014 and reviews PC's links to control theory (PID) and robotics. The authors argue PC's locality enables parallelizable training suited to neuromorphic hardware, while noting PC is not yet used industrially."},{"id":"arxiv:2410.02548","kind":"paper","n":81,"label":"Local Flow Matching Generative Models","authors":"Chen Xu, Xiuyuan Cheng, Yao Xie","year":2024,"url":"https://arxiv.org/abs/2410.02548","region":"t3","hemi":"wm","lobe":"occipital","color":"#9a714f","indegree":5,"alsoIn":[],"x":0.3663,"y":0.1457,"text":"Flow Matching (FM) trains a single global flow to interpolate between data and noise, which can require large models and long training. This paper introduces Local Flow Matching (LFM), which decomposes the global flow into a sequence of N small FM sub-models, each matching an Ornstein-Uhlenbeck diffusion step in the data-to-noise direction so that the interpolated distributions are closer together, enabling smaller sub-models and more efficient training. Leveraging the OU contraction and a \u03c7\u00b2-divergence analysis, the authors prove a generation guarantee of O(\u03b5^{1/2}) in \u03c7\u00b2-divergence (implying KL and TV bounds) with only N=O(log 1/\u03b5) blocks. Empirically LFM matches or beats global FM on tabular density estimation, image generation, and robotic manipulation policy learning, while training with less compute and being amenable to distillation."},{"id":"arxiv:2307.16125","kind":"paper","n":82,"label":"SEED-Bench: Benchmarking Multimodal LLMs with Generative Comprehension","authors":"Bohao Li et al.","year":2023,"url":"https://arxiv.org/abs/2307.16125","region":"t33","hemi":"llm","lobe":"parietal","color":"#ad67b0","indegree":4,"alsoIn":[],"x":0.8395,"y":0.1871,"text":"SEED-Bench is a large-scale benchmark for evaluating multimodal LLMs (MLLMs) on generative comprehension, consisting of 19K human-annotated multiple-choice questions across 12 evaluation dimensions spanning both image (spatial) and video (temporal) understanding. The authors develop a pipeline using foundation models for visual information extraction, ChatGPT/GPT-4 for question generation, automatic LLM-based filtering, and manual verification. They evaluate 18 models and find that most MLLMs exhibit limited performance, with InstructBLIP achieving the best overall accuracy of 53.37%, and VideoLLMs surprisingly failing to outperform ImageLLMs on temporal understanding tasks."},{"id":"arxiv:2510.20416","kind":"paper","n":83,"label":"Learning Coupled Earth System Dynamics with GraphDOP","authors":"Mihai Alexe, Eulalie Boucher, et al.","year":2025,"url":"https://arxiv.org/abs/2510.20416","region":"t34","hemi":"llm","lobe":"temporal","color":"#9a4f93","indegree":1,"alsoIn":[],"x":0.6121,"y":0.353,"text":"GraphDOP is a graph-based machine learning model that forecasts weather directly from raw satellite and in-situ observations, without using reanalysis products or physics-based NWP models, embedding diverse Earth System observations into a shared latent space so cross-component interactions emerge implicitly within a single model. The paper extends the original GraphDOP work with an autoregressive latent-space rollout and additional interface observations, then evaluates the model on three 2022 case studies where coupled processes dominate: rapid Arctic sea-ice refreezing, Hurricane Ian's ocean cold wake, and the European heatwave. GraphDOP reproduces sea-ice growth (comparable to ORAS6 reanalysis), storm track/structure and a sharp storm-induced SST cooling wake, and the northward advection of the heatwave, despite having no explicit ocean/sea-ice model coupling or sea-ice labels. Limitations include underestimated hurricane intensity (attributed to sparse in-situ pressure obs in the storm core) and damped heatwave extremes (linked to overestimated soil moisture)."},{"id":"arxiv:2602.10717","kind":"paper","n":84,"label":"Say, Dream, and Act: Learning Video World Models for Instruction-Driven Robot Manipulation","authors":"Tian et al.","year":2026,"url":"https://arxiv.org/abs/2602.10717","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":9,"alsoIn":[],"x":0.2948,"y":0.8538,"text":"The paper tackles instruction-driven robot manipulation by giving policies the ability to forecast how a scene evolves under actions. It proposes Dream4manip ('Say, Dream, Act'): select and domain-adapt a strong video generation backbone (Cosmos-Predict2) as a world model, apply latent-space adversarial distillation for few-step video prediction, and train an in-context conditioned action model that uses both imagined future frames and real observations to correct spatial errors. A length-agnostic imagination mechanism compresses arbitrary-length trajectories into a fixed set of keyframes for frame-rate-agnostic prediction. On the LIBERO benchmark the method reaches 98.1% total success with only a 1.22B backbone, outperforming larger VLA baselines."},{"id":"arxiv:2407.08867","kind":"paper","n":85,"label":"Perceptions of Sentient AI and Other Digital Minds: Evidence from the AI, Morality, and Sentience (AIMS) Survey","authors":"Jacy Reese Anthis, Janet V. T. Pauketat, Ali Ladak, Aikaterina Manoli","year":2024,"url":"https://arxiv.org/abs/2407.08867","region":"t7","hemi":"llm","lobe":"occipital","color":"#adb067","indegree":0,"alsoIn":[],"x":0.5822,"y":0.7265,"text":"This paper presents the first three waves (2021, 2023 main, and 2023 supplement; total N=3,500) of the nationally representative AI, Morality, and Sentience (AIMS) survey measuring U.S. adults' mind perception, moral concern, policy support, and forecasts regarding sentient AI and other 'digital minds.' Mind perception and moral concern for AI welfare were surprisingly high and significantly increased from 2021 to 2023, while opposition to building digital minds also rose. In 2023, ~one in five adults believed some current AI is sentient, 38% supported legal rights for sentient AI, and majorities supported banning sentient/smarter-than-human AI; the median forecast placed sentient AI just five years away. The authors argue safe and beneficial AI requires understanding public perceptions, with implications for XAI design, anthropomorphism tuning, and safety-focused policy."},{"id":"arxiv:2309.02591","kind":"paper","n":86,"label":"Scaling Autoregressive Multi-Modal Models: Pretraining and Instruction Tuning (CM3Leon)","authors":"Lili Yu, Bowen Shi, Ramakanth Pasunuru, et al. (Meta)","year":2023,"url":"https://arxiv.org/abs/2309.02591","region":"t23","hemi":"llm","lobe":"occipital","color":"#4f799a","indegree":0,"alsoIn":[],"x":0.8392,"y":0.2294,"text":"CM3Leon is a retrieval-augmented, token-based, decoder-only multi-modal language model that both generates and infills text and images, trained with a recipe adapted from text-only LLMs: large-scale retrieval-augmented pretraining on licensed Shutterstock data followed by multi-task supervised fine-tuning (SFT). It achieves state-of-the-art zero-shot text-to-image generation with 5x less training compute than comparable methods (zero-shot MS-COCO FID of 4.88), and introduces a self-contained contrastive decoding variant (CD-K) complementary to classifier-free guidance. After SFT, the model performs controllable tasks including text-guided image editing, structure/image-grounded generation, segmentation, and vision-language tasks like captioning and VQA. It shows autoregressive token models can rival and exceed diffusion models on cost-effectiveness and flexibility."},{"id":"arxiv:2503.06800","kind":"paper","n":87,"label":"VideoPhy-2: A Challenging Action-Centric Physical Commonsense Evaluation in Video Generation","authors":"Bansal, Peng, Lin et al.","year":2025,"url":"https://arxiv.org/abs/2503.06800","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":2,"alsoIn":[],"x":0.4466,"y":0.8313,"text":"VideoPhy-2 is an action-centric benchmark for evaluating physical commonsense in text-to-video generation, built from 197 real-world actions (sports/physical activities and object interactions), 3940 LLM-generated prompts, 6800 generated videos, and 102K human annotations rating semantic adherence (SA), physical commonsense (PC), and fine-grained physical-rule violations on a 5-point scale. Evaluating five open and two closed video models, even the best model (Wan2.1-14B) reaches only 32.6% joint (SA\u22654 and PC\u22654) on the full set and 21.9% on a hard subset, with conservation of mass and momentum the most-violated laws (~40%). The authors also train VideoPhy-2-AutoEval, a 7B multi-task automatic evaluator fine-tuned on ~50K annotations that substantially outperforms Gemini-2.0-Flash-Exp. Overall the benchmark exposes large gaps in video models as physical world simulators."},{"id":"arxiv:2510.03847","kind":"paper","n":88,"label":"Small Language Models for Agentic Systems: A Survey of Architectures, Capabilities, and Deployment Trade offs","authors":"","year":null,"url":"https://arxiv.org/abs/2510.03847","region":"t35","hemi":"llm","lobe":"occipital","color":"#b0679e","indegree":2,"alsoIn":[],"x":0.8649,"y":0.7785,"text":"This survey argues that Small Language Models (SLMs, 1\u201312B params) are sufficient and often superior to frontier LLMs for agentic workloads such as function calling, structured decoding, and RAG, due to schema-constrained accuracy dominating open-ended generalization. It synthesizes results across open and proprietary SLMs (Phi-4-Mini, Qwen-2.5-7B, Gemma-2-9B, Llama-3.2, Ministral, DeepSeek-R1-Distill, Apple on-device), connects them to benchmarks (BFCL v4, StableToolBench) and serving stacks (vLLM/SGLang/TensorRT-LLM + XGrammar/Outlines), and formalizes an SLM-default, LLM-fallback architecture with uncertainty-aware routing. The paper proposes engineering metrics (Cost per Successful task, schema validity, executable-call rate) and reports that guided decoding lets SLMs match or surpass LLMs at 10\u00d7\u2013100\u00d7 lower token cost."},{"id":"openreview.net:c15700430713","kind":"paper","n":89,"label":"WorldSplat: 4D-aware controllable driving scene world model","authors":"(ICLR 2026 authors)","year":2026,"url":"https://openreview.net/pdf/26fbb3a9ef84175c8a2efe7918a32cd5a0082627.pdf","region":"t15","hemi":"wm","lobe":"occipital","color":"#4f9a62","indegree":3,"alsoIn":[],"x":0.444,"y":0.6548,"text":"WorldSplat addresses the dilemma between driving-scene video generation (high fidelity but weak 3D consistency and poor novel-view support) and 3D/4D reconstruction (good novel-view synthesis but no generative ability) by proposing a feed-forward framework for 4D driving-scene generation. It uses a 4D-aware latent diffusion model that produces multi-modal latents (RGB, metric depth, semantic masks) from user conditions, a latent Gaussian decoder that converts them into pixel-aligned 3D Gaussians decomposed into static/dynamic parts and aggregated into a 4D representation, and an enhanced video diffusion model that refines rendered novel-view frames. On nuScenes it achieves state-of-the-art video generation (e.g. 74.13 FVD, 8.78 FID without first-frame guidance) and novel-view synthesis under lateral ego shifts, outperforming reconstruction baselines like OmniRe and DiST-4D, while also improving downstream perception training."},{"id":"arxiv:2207.12316","kind":"paper","n":90,"label":"A Theoretical Framework for Inference and Learning in Predictive Coding Networks","authors":"Beren Millidge, Yuhang Song, Tommaso Salvatori, Thomas Lukasiewicz, Rafal Bogacz","year":2022,"url":"https://arxiv.org/abs/2207.12316","region":"t20","hemi":"llm","lobe":"frontal","color":"#67b0b0","indegree":4,"alsoIn":[],"x":0.6331,"y":0.6408,"text":"Predictive coding networks (PCNs) trained with prospective configuration match backpropagation (BP) empirically but lacked theoretical grounding; this paper provides a comprehensive theoretical analysis of both the inference and learning phases. For inference, it derives an analytical linear equilibrium showing PCN equilibrium activities interpolate between BP gradients (feedforward-dominant) and target propagation (TP) targets (feedback-dominant), controlled by feedback/feedforward precision ratios. For learning, it reinterprets PC as a generalized expectation-maximization algorithm with a constrained E-step and proves that PCNs converge to critical points of the BP loss, establishing that deep PCNs have in-principle equal learning capacity to BP while retaining local Hebbian updates. Empirical results on MNIST confirm the theory (e.g. BP gradient norm ~1e-16 for a single digit, equivalent accuracy to BP)."},{"id":"arxiv:2303.01469","kind":"paper","n":91,"label":"Consistency Models","authors":"Yang Song, Prafulla Dhariwal, Mark Chen, Ilya Sutskever","year":2023,"url":"https://arxiv.org/abs/2303.01469","region":"t32","hemi":"llm","lobe":"frontal","color":"#8b4f9a","indegree":4,"alsoIn":[],"x":0.6749,"y":0.2667,"text":"Diffusion models produce high-quality images but require slow iterative sampling. This paper introduces consistency models, a new family of generative models that learn to map any point on a Probability Flow ODE trajectory directly back to its origin (the data sample), enabling single-step generation while still supporting multistep sampling to trade compute for quality and zero-shot data editing. They can be trained either by distilling a pre-trained diffusion model (consistency distillation, CD) or standalone from scratch (consistency training, CT), neither requiring adversarial training. CD sets new state-of-the-art one-step FID of 3.55 on CIFAR-10 and 6.20 on ImageNet 64x64, and as a standalone model CT outperforms prior single-step non-adversarial generators."},{"id":"transformer-circuits.pub:e2306116790e","kind":"paper","n":92,"label":"Towards Monosemanticity: Decomposing Language Models With Dictionary Learning","authors":"Trenton Bricken, Adly Templeton, Joshua Batson, et al. (Anthropic)","year":2023,"url":"https://transformer-circuits.pub/2023/monosemantic-features/index.html","region":"t16","hemi":"llm","lobe":"frontal","color":"#67b084","indegree":5,"alsoIn":[],"x":0.6615,"y":0.4076,"text":"The paper tackles polysemanticity in neural networks \u2014 the fact that individual neurons respond to mixtures of unrelated inputs, hypothesized to arise from superposition (representing more features than neurons). The authors train a sparse autoencoder (ReLU encoder + linear decoder, L1 sparsity penalty plus MSE reconstruction) on the 512-neuron MLP activations of a one-layer transformer, decomposing them into an overcomplete dictionary of up to 131,072 features. They show via detailed case studies (Arabic, DNA, base64, Hebrew features), human scoring, and automated interpretability that the learned features are far more monosemantic and interpretable than neurons, are causally used by the model, are largely universal across independently trained models, and exhibit 'feature splitting' as dictionary size grows. The focus run A/1 (4,096 features) recovers 79% of the MLP's log-likelihood loss contribution."},{"id":"arxiv:1908.05656","kind":"paper","n":93,"label":"PHYRE: A New Benchmark for Physical Reasoning","authors":"Bakhtin, van der Maaten, Johnson, Gustafson, Girshick","year":2019,"url":"https://arxiv.org/abs/1908.05656","region":"t6","hemi":"wm","lobe":"temporal","color":"#b0a967","indegree":0,"alsoIn":[],"x":0.4362,"y":0.8073,"text":"The paper introduces PHYRE, a benchmark of 2D classical-mechanics puzzles where an agent must place new dynamic bodies (balls) into a scene so that, after deterministic physics simulation, a goal relation (one body touching another for \u22653 seconds) is satisfied. It is designed to test physical reasoning, generalization (within-template vs cross-template), and sample-efficiency, with two tiers (PHYRE-B: one ball, 3D action space; PHYRE-2B: two balls, 6D), each with 25 task templates \u00d7 100 tasks. The authors benchmark five agents (random, non-parametric memory, DQN, and online variants) using an AUCCESS metric that weights solving in few attempts. They find modern methods generalize poorly, especially cross-template on PHYRE-2B (best agent only 39.6% AUCCESS), motivating research on more sample-efficient physical-reasoning agents."},{"id":"arxiv:2410.18072","kind":"paper","n":94,"label":"WorldSimBench: Towards Video Generation Models as World Simulators","authors":"Qin, Shi, Yu et al.","year":2024,"url":"https://arxiv.org/abs/2410.18072","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":6,"alsoIn":[],"x":0.3792,"y":0.8916,"text":"WorldSimBench addresses the lack of an embodied-perspective evaluation for high-capability predictive ('World Simulator') video models by first organizing predictive models into an S0\u2013S3 capability hierarchy, then proposing a dual evaluation framework. Explicit Perceptual Evaluation trains a Human Preference Evaluator (LoRA-fine-tuned Flash-VStream VideoLLM) on the new HF-Embodied Dataset (35,701 human-feedback tuples) to score visual fidelity, while Implicit Manipulative Evaluation closes the loop by converting generated videos into actions via pre-trained video-to-action policies in MineRL, CARLA, and CALVIN. Evaluating 8 video generation models across Open-Ended Environment, Autonomous Driving, and Robot Manipulation, the Human Preference Evaluator aligns far better with humans than GPT-4o (e.g., OE accuracy 89.4 vs 72.8). The authors conclude current video generators still fail to capture many physical rules and are not yet reliable World Simulators."},{"id":"cambridge.org:57359c90d672","kind":"paper","n":95,"label":"The best game in town: the reemergence of the language-of-thought hypothesis across the cognitive sciences","authors":"Quilty-Dunn, J., Porot, N. & Mandelbaum, E.","year":2023,"url":"https://www.cambridge.org/core/journals/behavioral-and-brain-sciences/article/abs/best-game-in-town-the-reemergence-of-the-languageofthought-hypothesis-across-the-cognitive-sciences/76F46784C6C07FF52FF45B934D6D3542","region":"t5","hemi":"llm","lobe":"parietal","color":"#b09e67","indegree":0,"alsoIn":[],"x":0.7333,"y":0.7885,"text":"This target article argues that the language-of-thought (LoT) hypothesis\u2014classical compositional, symbolic mental representation\u2014has reemerged as the best-supported account of representational format across cognitive science, rather than being an obsolete relic. The authors define LoT by six clustering core properties and survey converging evidence from computational, perceptual, developmental, comparative, and social psychology. They conclude that diverse phenomena\u2014probabilistic/Bayesian program-induction models, compositional object perception, infant and animal logical reasoning, and automatic intuitive adult cognition\u2014all implicate LoT-like structures. They grant representational pluralism (iconic, associative, and deep-neural-network-like formats coexist) but maintain that classical symbolic LoT architectures remain uniquely flexible and explanatorily broad."},{"id":"nature.com:7719482cd2c4","kind":"paper","n":96,"label":"Microstructure of a spatial map in the entorhinal cortex","authors":"Hafting, T., Fyhn, M., Molden, S., Moser, M.-B. & Moser, E. I.","year":2005,"url":"https://www.nature.com/articles/nature03721","region":"t28","hemi":"llm","lobe":"frontal","color":"#7667b0","indegree":2,"alsoIn":[],"x":0.5807,"y":0.3351,"text":"The paper investigates how the brain integrates information about place, distance and direction during navigation, an operation whose cortical implementation was poorly understood. By recording neurons in the dorsocaudal medial entorhinal cortex (dMEC) of freely moving rats, the authors discover 'grid cells' that fire whenever the animal occupies any vertex of a regular triangular (hexagonal) grid tiling the environment. Neighbouring grid cells share grid orientation and spacing but differ in spatial phase, grid spacing and field size increase along the dorsal-to-ventral axis, and the grid is anchored to landmarks yet persists in darkness, implying a path-integration-based metric map of space."},{"id":"transformer-circuits.pub:c4d9a1f0e2b7","kind":"paper","n":97,"label":"Verbalizable Representations Form a Global Workspace in Language Models","authors":"W. Gurnee, N. Sofroniew, A. Pearce, J. Lindsey, et al. (Anthropic)","year":2026,"url":"https://transformer-circuits.pub/2026/workspace/index.html","region":"t6","hemi":"llm","lobe":"temporal","color":"#b0a967","indegree":1,"alsoIn":[],"x":0.648,"y":0.4471,"text":"Anthropic interpretability paper (Gurnee, Sofroniew, Lindsey et al., July 6 2026) showing that language models maintain a workspace-like set of internal representations analogous to conscious access in Global Workspace Theory. It introduces the Jacobian lens (J-lens), which finds the internal vectors a model is poised to verbalize by averaging the linearized causal effect of a middle-layer activation on the model's next-token likelihoods over about 1,000 prompts, correcting the older logit lens so that early-layer content becomes legible. Across five functional properties the workspace representations can be reported, deliberately summoned, used as causally load-bearing unspoken intermediate steps in reasoning, routed to many downstream operations, and are selectively required for flexible reasoning but not for automatic tasks. The workspace occupies roughly the middle-to-late band of model depth (about 38 to 92 percent), with early layers input-adjacent and late layers output-adjacent, mirroring a perceptual, conceptual, then motor hierarchy. The paper is careful about scope: it takes no position on subjective experience, notes the workspace holds only about 10 to 20 concepts at a time and is rebuilt each forward pass with no persistent state, and offers the J-lens as a safety-auditing tool that can surface hidden reasoning such as manipulation or evaluation-awareness not present in the output."},{"id":"arxiv:2308.03688","kind":"paper","n":98,"label":"AgentBench: Evaluating LLMs as Agents","authors":"Xiao Liu et al.","year":2023,"url":"https://arxiv.org/abs/2308.03688","region":"t35","hemi":"llm","lobe":"occipital","color":"#b0679e","indegree":12,"alsoIn":[],"x":0.6306,"y":0.9491,"text":"The paper introduces AgentBench, a multi-dimensional benchmark of 8 distinct interactive environments (Operating System, Database, Knowledge Graph, Digital Card Game, Lateral Thinking Puzzles, House-Holding/ALFWorld, Web Shopping/WebShop, Web Browsing/Mind2Web) to evaluate LLMs as autonomous agents on reasoning and decision-making via Chain-of-Thought prompting. It evaluates 29 API-based and open-source LLMs using a unified server-client toolkit, finding top commercial models (gpt-4 overall 4.01) vastly outperform OSS models \u226470B (codellama-34b best OSS at 0.96), with an average gap of 2.32 vs 0.51. Failure analysis attributes most agent failures to Task Limit Exceeded, plus poor instruction following (Invalid Format/Action), pointing to weak long-term reasoning and decision-making. It also finds code training is a double-edged sword and high-quality multi-round alignment data improves agent performance."},{"id":"arxiv:2309.17080","kind":"paper","n":99,"label":"GAIA-1: A Generative World Model for Autonomous Driving","authors":"Hu et al. (Wayve)","year":2023,"url":"https://arxiv.org/abs/2309.17080","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":2,"alsoIn":[],"x":0.3222,"y":0.7456,"text":"GAIA-1 is a generative world model for autonomous driving that ingests video, text, and action inputs and casts world modeling as next-token prediction over discrete tokens. It pairs a 6.5B-parameter autoregressive transformer world model (over VQ image tokens with DINO-distilled semantics) with a 2.6B-parameter video diffusion decoder that renders predicted tokens back to high-resolution, temporally-consistent video and temporally upsamples to 25Hz. Trained on 4,700 hours (~420M images) of London urban driving data, it generates realistic, controllable driving scenarios, exhibits emergent properties (3D geometry, reactive agents, extrapolation to out-of-distribution behaviors like steering off-lane), and follows LLM-style scaling laws where final performance was predicted from models <20x smaller in compute."},{"id":"nature.com:7fc0dd8e4038","kind":"paper","n":100,"label":"The free-energy principle: a unified brain theory?","authors":"Friston, K.","year":2010,"url":"https://www.nature.com/articles/nrn2787","region":"t25","hemi":"llm","lobe":"parietal","color":"#6779b0","indegree":1,"alsoIn":[],"x":0.5673,"y":0.8404,"text":"Karl Friston's review proposes the free-energy principle as a single optimization framework that unifies action, perception and learning in the brain. It argues that adaptive agents resist disorder by minimizing a free-energy bound on surprise (surprisal) \u2014 perception minimizes free energy with respect to neuronal activity, synaptic efficacy and gain, while action minimizes it by changing sensory samples. The review then maps major brain theories (the Bayesian brain, predictive coding, the infomax/efficient-coding principle, neural Darwinism, optimal control and reinforcement learning) onto this single quantity, showing each optimizes value (expected reward) or its complement, surprise. Simulations of birdsong categorization, cued reaching and the mountain-car problem illustrate perception and action as free-energy minimization under generative models."},{"id":"arxiv:2606.24467","kind":"paper","n":101,"label":"CompressKV: Semantic-Retrieval-Guided KV-Cache Compression for Resource-Efficient Long-Context LLM Inference","authors":"","year":null,"url":"https://arxiv.org/abs/2606.24467","region":"t18","hemi":"llm","lobe":"temporal","color":"#67b09a","indegree":6,"alsoIn":[],"x":0.8783,"y":0.4116,"text":"Long-context LLM inference is bottlenecked by KV-cache memory, and existing eviction methods ignore attention head heterogeneity in GQA-based models, allowing Streaming Heads to dominate token eviction and discard critical mid-context tokens. CompressKV identifies Semantic Retrieval Heads (SRHs)\u2014heads that attend to entire answer spans rather than just top-k tokens\u2014to guide token retention, and allocates per-layer cache budgets proportionally to offline-estimated Frobenius-norm eviction errors. Evaluated on LongBench and Needle-in-a-Haystack across four LLMs, CompressKV consistently outperforms six baselines (StreamingLLM, SnapKV, PyramidKV, CAKE, HeadKV, AdaKV), especially under tight memory budgets."},{"id":"arxiv:2604.06882","kind":"paper","n":102,"label":"Telecom World Models: Unifying Digital Twins, Foundation Models, and Predictive Planning for 6G","authors":"Hang Zou, Yuzhi Yang, Lina Bariah, et al.","year":2026,"url":"https://arxiv.org/abs/2604.06882","region":"t39","hemi":"wm","lobe":"occipital","color":"#9a4f5b","indegree":5,"alsoIn":[],"x":0.2056,"y":0.5089,"text":"The paper introduces the Telecom World Model (TWM), a learned, action-conditioned, uncertainty-aware architecture for 6G network management that unifies the strengths of LLMs/foundation models (flexible reasoning, no network dynamics) and Digital Twins (high-fidelity simulation, but not learned or planning-oriented). TWM is formalized as a factored-state POMDP that splits the telecom state into a Controllable System World (operator configs/policies) and an Exogenous World (propagation, mobility, traffic, failures), and is realized as three layers: a Field World Model (FWM) for spatial/EM field prediction, a Control/Dynamics World Model (CDWM) for action-conditioned KPI trajectory rollout, and a TelecomGPT foundation-model layer for intent translation, orchestration, and guardrails. A qualitative comparison argues TWM is the first construct to jointly provide telecom state grounding, fast rollouts, calibrated uncertainty, multi-timescale dynamics, model-based planning, and LLM guardrails. A network-slicing proof-of-concept shows the full three-layer pipeline reaches the cost\u2013compliance Pareto front (avg SLA Compliance Margin 45%, ~7\u20138% cost reduction), outperforming single-world baselines and a budget-matched Digital-Twin search."},{"id":"arxiv:1803.10122","kind":"paper","n":103,"label":"Recurrent World Models Facilitate Policy Evolution (World Models)","authors":"Ha & Schmidhuber","year":2018,"url":"https://arxiv.org/abs/1803.10122","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":7,"alsoIn":[],"x":0.0837,"y":0.4417,"text":"The paper proposes 'World Models,' a framework that splits an RL agent into a large unsupervised world model and a tiny controller: a Variational Autoencoder (V) compresses each image frame into a latent vector z, an MDN-RNN (M) predicts the distribution of the next latent given the action and history, and a small linear controller (C) maps [z, h] to actions, optimized with CMA-ES. Trained from random rollouts, this approach solves the CarRacing-v0 task with a score of 906\u00b121, exceeding prior Deep RL and leaderboard results. The authors further show an agent can be trained entirely inside the M-generated 'dream' of a VizDoom Take Cover environment and transferred back to the real environment (~1100 time steps), and that raising the MDN-RNN's temperature \u03c4 counters the controller's tendency to exploit imperfections in the learned model."},{"id":"arxiv:2502.11831","kind":"paper","n":104,"label":"Intuitive physics understanding emerges from self-supervised pretraining on natural videos","authors":"Quentin Garrido, Nicolas Ballas, Mahmoud Assran, Adrien Bardes, Laurent Najman, ","year":2025,"url":"https://arxiv.org/abs/2502.11831","region":"t20","hemi":"wm","lobe":"frontal","color":"#67b0b0","indegree":4,"alsoIn":[],"x":0.2678,"y":0.5152,"text":"The paper tests whether intuitive physics understanding (object permanence, continuity, shape constancy, etc.) emerges in deep networks trained only to predict masked regions of natural videos, using the developmental-psychology violation-of-expectation paradigm where a model's prediction error ('surprise') flags physically impossible events. V-JEPA, which predicts masked video content in a learned representation space, distinguishes possible from impossible videos far above chance (98% on IntPhys, 66% on GRASP, 62% on InfLevel-lab), whereas pixel-prediction (VideoMAEv2) and text-reasoning MLLMs (Qwen2-VL, Gemini 1.5 pro) perform near chance. The effect is robust across masking strategy, model size (even 115M params), and data scale (even 128h of unique video), with V-JEPA matching or exceeding human accuracy on the IntPhys test set. The authors conclude that prediction in an abstract representation space \u2014 akin to predictive coding \u2014 is sufficient for intuitive physics, challenging the view that hardwired 'core knowledge' is required."},{"id":"arxiv:2406.14798","kind":"paper","n":105,"label":"Probabilistic Emulation of a Global Climate Model with Spherical DYffusion","authors":"Salva Ruhling Cachay, Brian Henn, Oliver Watt-Meyer, Christopher S. Bretherton, ","year":2024,"url":"https://arxiv.org/abs/2406.14798","region":"t34","hemi":"wm","lobe":"temporal","color":"#9a4f93","indegree":2,"alsoIn":[],"x":0.2704,"y":0.1339,"text":"Data-driven weather forecasting models do not directly transfer to climate modeling, which requires stable multi-decade simulations and accurate reproduction of long-term statistics. The authors propose Spherical DYffusion, which integrates the dynamics-informed diffusion framework (DYffusion) with the Spherical Fourier Neural Operator (SFNO) architecture, adding time-conditioning and inference stochasticity modules to enable probabilistic ensemble climate emulation of the FV3GFS physics-based model. The method achieves stable 100-year simulations at 6-hourly timesteps with climate biases within 50% of the reference noise floor on average across 34 fields, more than 2x lower than the next best baseline (ACE), while requiring less than 3x the compute of a deterministic model."},{"id":"arxiv:1312.6114","kind":"paper","n":106,"label":"Auto-Encoding Variational Bayes","authors":"Diederik P. Kingma, Max Welling","year":2013,"url":"https://arxiv.org/abs/1312.6114","region":"t3","hemi":"llm","lobe":"occipital","color":"#9a714f","indegree":1,"alsoIn":[],"x":0.648,"y":0.5663,"text":"The paper tackles efficient approximate inference and learning in directed probabilistic models that have continuous latent variables with intractable posteriors, scaled to large datasets. It introduces the SGVB estimator, which uses a reparameterization of the variational lower bound (writing z = g_\u03c6(\u03b5,x) with \u03b5~p(\u03b5)) to produce a low-variance, differentiable, unbiased Monte Carlo estimator optimizable by standard SGD, and the AEVB algorithm, which fits a neural recognition model (encoder) jointly with the generative model (decoder) \u2014 yielding the variational auto-encoder. Experiments on MNIST and Frey Face show AEVB converges faster and to a better variational lower bound than the wake-sleep algorithm across all latent dimensionalities, and beats wake-sleep and Monte Carlo EM on estimated marginal likelihood."},{"id":"arxiv:2312.03788","kind":"paper","n":107,"label":"SmoothQuant+: Accurate and Efficient 4-bit Post-Training WeightQuantization for LLM","authors":"","year":null,"url":"https://arxiv.org/abs/2312.03788","region":"t37","hemi":"llm","lobe":"parietal","color":"#b06788","indegree":4,"alsoIn":[],"x":0.875,"y":0.3933,"text":"Addresses accuracy loss in 4-bit weight-only post-training quantization (PTQ) of LLMs, attributing the loss to weight quantization error being amplified by activation outliers. SmoothQuant+ applies a mathematically-equivalent per-channel smoothing transform (controlled by strength \u03b1, fused into preceding layers) to suppress activation outliers and adjust weights before group-wise (group-size 128) 4-bit weight quantization, with \u03b1 chosen via grid search minimizing whole-model quantization loss. Integrated into vLLM with an efficient W4A16 CUDA kernel adapted from LMDeploy. On the Code Llama family it achieves lossless (even improved) HumanEval accuracy, letting Code Llama-34B run on a single A100 40GB GPU with 1.9\u20134.0\u00d7 throughput and 68% per-token latency versus FP16 on two A100 40GB GPUs."},{"id":"arxiv:2407.00118","kind":"paper","n":108,"label":"From Efficient Multimodal Models to World Models: A Survey","authors":"Xinji Mai, Zeng Tao, Junxiong Lin, Haoran Wang, Yang Chang, Yanlan Kang, Yan Wan","year":2024,"url":"https://arxiv.org/abs/2407.00118","region":"t17","hemi":"llm","lobe":"parietal","color":"#4f9a79","indegree":4,"alsoIn":[],"x":0.7303,"y":0.2976,"text":"This survey reviews the development of efficient Multimodal Large Models (MLMs) and frames them as a pathway toward world models capable of simulating and predicting environmental state changes. It surveys foundational architectures (Transformers and their efficient challengers like linear attention, RWKV, Mamba/Mamba-2, JEPA), model-optimization techniques (compression, pruning, knowledge distillation, quantization, fine-tuning), and multimodal-specific techniques (M-IT, M-ICL, M-COT, encoders, generative models). It contrasts the two main routes to world models \u2014 OpenAI's autoregressive/data-driven approach (e.g. Sora) versus Meta's JEPA hierarchical-planning approach \u2014 and proposes augmenting MLMs with 3D generation, embodied intelligence, and external rule systems to enhance world simulation and counterfactual reasoning."},{"id":"arxiv:1709.06560","kind":"paper","n":109,"label":"Deep Reinforcement Learning that Matters","authors":"Henderson et al.","year":2017,"url":"https://arxiv.org/abs/1709.06560","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":5,"alsoIn":[],"x":0.1881,"y":0.2549,"text":"This paper investigates the reproducibility of state-of-the-art deep reinforcement learning, focusing on model-free policy gradient methods (TRPO, DDPG, PPO, ACKTR) on MuJoCo continuous-control tasks from OpenAI Gym. Through controlled experiments it shows that reported results are highly sensitive to extrinsic factors (hyperparameters such as network architecture, activation functions, reward scaling, batch size, and choice of codebase) and intrinsic factors (random seeds, environment dynamics), making fair comparison difficult. It demonstrates that even identical hyperparameters can yield statistically different performance distributions purely from random seeds, and that implementation differences not described in papers dramatically change results. It recommends running many seeds, reporting all settings/code, and applying significance testing (bootstrap confidence intervals, t-tests, KS tests, power analysis)."},{"id":"arxiv:2204.00227","kind":"paper","n":110,"label":"Perception Prioritized Training of Diffusion Models","authors":"Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, Sungroh Yoo","year":2022,"url":"https://arxiv.org/abs/2204.00227","region":"t32","hemi":"wm","lobe":"frontal","color":"#8b4f9a","indegree":4,"alsoIn":[],"x":0.374,"y":0.1085,"text":"The paper investigates what a diffusion model learns at each noise level and shows that recovering images corrupted at certain (medium-SNR) noise levels is the pretext task through which the model learns perceptually rich visual content, while high-SNR levels only teach imperceptible details. It proposes P2 (Perception Prioritized) weighting, a redesign of the denoising-score-matching loss weighting \u03bb'_t = \u03bb_t/(k+SNR(t))^\u03b3 that down-weights the 'clean-up' stage and emphasizes the 'content' stage. Across datasets, architectures, and sampling step counts, P2 weighting consistently improves FID/KID over the standard (Ho et al.) weighting baseline. It achieves state-of-the-art FID on CelebA-HQ and Oxford Flowers and competitive FID on FFHQ among diverse generative model families."},{"id":"arxiv:2305.10626","kind":"paper","n":111,"label":"Language Models Meet World Models: Embodied Experiences Enhance Language Models","authors":"Xiang et al.","year":2023,"url":"https://arxiv.org/abs/2305.10626","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":10,"alsoIn":[],"x":0.6869,"y":0.835,"text":"Pretrained language models lack embodied knowledge (object permanence, household planning) because they train only on text. The paper proposes E2WM, a paradigm that finetunes LMs on embodied experiences collected from a world-model simulator (VirtualHome) via goal-oriented planning (using MCTS) and random exploration, compiled into plan-generation, activity-recognition, counting, and object-path-tracking tasks. To preserve general language ability and efficiency, it introduces EWC-LoRA, combining elastic weight consolidation with low-rank adapters. The approach improves base LMs (1.3B\u201313B) by 64.28% on average across 18 downstream tasks, with small finetuned models matching or beating ChatGPT while keeping Pile perplexity nearly unchanged."},{"id":"arxiv:2506.09985","kind":"paper","n":112,"label":"V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning","authors":"Assran, Bardes, Fan, Garrido et al. (Meta FAIR)","year":2025,"url":"https://arxiv.org/abs/2506.09985","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":3,"alsoIn":[],"x":0.2833,"y":0.4774,"text":"V-JEPA 2 is a self-supervised joint-embedding-predictive video model pretrained on over 1 million hours of internet video (22M-sample VideoMix22M) with a mask-denoising feature-prediction objective, scaled to a 1B-parameter ViT-g via data scaling, model scaling, longer training, and progressive resolution. It achieves state-of-the-art motion understanding and human action anticipation, and when aligned with an LLM reaches SOTA video QA at the 8B scale. The authors then post-train an action-conditioned world model (V-JEPA 2-AC) on under 62 hours of unlabeled Droid robot video, freezing the encoder and learning a 300M block-causal predictor, and deploy it zero-shot for planning-based pick-and-place on Franka arms in new labs without task-specific training or reward."},{"id":"arxiv:2110.04544","kind":"paper","n":113,"label":"CLIP-Adapter: Better Vision-Language Models with Feature Adapters","authors":"","year":null,"url":"https://arxiv.org/abs/2110.04544","region":"t23","hemi":"wm","lobe":"occipital","color":"#4f799a","indegree":6,"alsoIn":[],"x":0.33,"y":0.3753,"text":"Few-shot adaptation of the pretrained CLIP vision-language model normally relies on prompt tuning (e.g. CoOp), which learns continuous soft prompts but is slow and memory-heavy due to back-propagation through the text encoder. This paper proposes CLIP-Adapter, which freezes CLIP and appends a lightweight two-layer bottleneck adapter to the visual (or text) branch, then blends the adapted feature with the original CLIP feature via a residual connection controlled by a residual ratio. With only 0.52M trainable parameters and hand-crafted hard prompts, CLIP-Adapter outperforms zero-shot CLIP, linear-probe CLIP and CoOp on all 11 classification datasets and few-shot settings while training far faster. On 16-shot ImageNet it reaches 61.33% vs CoOp's 60.46% using 50min training (vs CoOp's 14h40min) and 2227 MiB GPU memory (vs 7193 MiB)."},{"id":"arxiv:2401.15884","kind":"paper","n":114,"label":"Corrective Retrieval Augmented Generation","authors":"Shi-Qi Yan, Jia-Chen Gu, Yun Zhu, Zhen-Hua Ling","year":2024,"url":"https://arxiv.org/abs/2401.15884","region":"t31","hemi":"llm","lobe":"occipital","color":"#9767b0","indegree":1,"alsoIn":[],"x":0.8544,"y":0.6019,"text":"RAG quality degrades badly when retrieval returns irrelevant documents, since generators indiscriminately consume whatever is retrieved. This paper proposes CRAG, which adds a lightweight T5-large (0.77B) retrieval evaluator that scores document relevance and triggers one of three actions \u2014 Correct (refine via decompose-then-recompose), Incorrect (discard and fall back to web search), or Ambiguous (combine both) \u2014 before generation. CRAG is plug-and-play atop standard RAG and Self-RAG, and across four datasets (PopQA, Biography, PubHealth, Arc-Challenge) it consistently and significantly improves accuracy/FactScore, e.g. Self-CRAG beats Self-RAG by 20.0% accuracy on PopQA with LLaMA2-hf-7b. It also improves robustness to degrading retrieval quality at only modest computational overhead."},{"id":"arxiv:2501.09038","kind":"paper","n":115,"label":"Do generative video models understand physical principles?","authors":"Motamed, Culp, Swersky, Jaini, Geirhos et al. (INSAIT & Google DeepMind)","year":2025,"url":"https://arxiv.org/abs/2501.09038","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":8,"alsoIn":[],"x":0.4415,"y":0.7726,"text":"The paper asks whether generative video models acquire physical understanding or merely produce visually realistic pixels. The authors introduce Physics-IQ, a real-world benchmark of 396 videos (66 scenarios \u00d7 3 camera views \u00d7 2 takes, 8s each) spanning solid mechanics, fluid dynamics, optics, thermodynamics and magnetism, where models must predict 5 seconds of continuation from conditioning frames; predictions are scored against ground truth via four motion-based metrics (Spatial IoU, Spatiotemporal IoU, Weighted Spatial IoU, MSE) combined into a Physics-IQ score normalized so real-vs-real physical variance = 100%. Across eight model variants (Sora, Runway Gen 3, Pika 1.0, Lumiere, Stable Video Diffusion, VideoPoet), physical understanding is severely limited \u2014 the best model, VideoPoet (multiframe), reaches only 29.5%. Crucially, visual realism (measured by a Gemini-1.5-Pro 2AFC test) is uncorrelated with physical understanding (Pearson r = -0.46, p=.249), showing realism does not imply physics understanding."},{"id":"arxiv:2502.05855","kind":"paper","n":116,"label":"DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control","authors":"Junjie Wen, Yichen Zhu, Jinming Li, et al.","year":2025,"url":"https://arxiv.org/abs/2502.05855","region":"t0","hemi":"wm","lobe":"frontal","color":"#9a4f4f","indegree":4,"alsoIn":[],"x":0.1738,"y":0.7618,"text":"DexVLA addresses the action representation bottleneck in VLA models by introducing a 1B-parameter diffusion-based action expert with multi-head architecture for cross-embodiment learning, combined with a three-stage embodied curriculum (cross-embodiment pre-training, embodiment-specific alignment, task-specific adaptation). The model uses substep reasoning\u2014generating intermediate language sub-instructions\u2014to handle long-horizon tasks without external high-level planners like SayCan. Evaluated across single-arm, bimanual, and dexterous hand embodiments on tasks ranging from shirt folding to laundry folding, DexVLA outperforms OpenVLA, Octo, Diffusion Policy, and \u03c00 while being pre-trained on only 100 hours of data and running at 60Hz on a single A6000 GPU."},{"id":"arxiv:2310.08560","kind":"paper","n":117,"label":"MemGPT: Towards LLMs as Operating Systems","authors":"Packer et al.","year":2023,"url":"https://arxiv.org/abs/2310.08560","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":1,"alsoIn":[],"x":0.6494,"y":0.9321,"text":"MemGPT tackles the fixed-context-window limitation of LLMs that hampers extended conversations and long-document reasoning. Borrowing from OS virtual memory/paging, it introduces 'virtual context management': a tiered memory hierarchy (main context = prompt tokens, external context = recall/archival storage) where the LLM uses self-directed function calls to page data in and out of its own context, plus a queue manager that handles eviction and recursive summarization under 'memory pressure'. Evaluated on conversational agents (Deep Memory Retrieval, Conversation Opener) and document analysis (multi-document QA, nested key-value retrieval), MemGPT substantially outperforms fixed-context baselines \u2014 e.g. lifting GPT-4 DMR accuracy from 32.1% to 92.5% \u2014 and is the only method to handle nested KV retrieval beyond 2 nesting levels."},{"id":"arxiv:2305.15074","kind":"paper","n":118,"label":"Have LLMs Advanced Enough? A Challenging Problem Solving Benchmark (JEEBench)","authors":"Daman Arora et al.","year":2023,"url":"https://arxiv.org/abs/2305.15074","region":"t33","hemi":"llm","lobe":"parietal","color":"#ad67b0","indegree":3,"alsoIn":[],"x":0.7991,"y":0.5956,"text":"The paper introduces JEEBench, a benchmark of 515 hard pre-engineering Physics, Chemistry and Mathematics problems curated from 8 editions (2016\u20132023) of India's IIT JEE-Advanced exam, requiring long-horizon reasoning over deep domain knowledge. Evaluating open-source and proprietary LLMs, the best model (GPT-4) reaches under 40% even with chain-of-thought, self-consistency and self-refinement. Error analysis attributes GPT-4's failures mainly to conceptual retrieval errors and algebraic/arithmetic computation errors, and shows self-critique does not help. The authors add a post-hoc confidence-thresholding method over self-consistency to handle negative-marking risk, placing GPT-4 in the top 10\u201320 percentile of human exam-takers in 2023."},{"id":"arxiv:2602.24044","kind":"paper","n":119,"label":"Data Driven Optimization of GPU efficiency for Distributed LLM Adapter Serving","authors":"","year":null,"url":"https://arxiv.org/abs/2602.24044","region":"t24","hemi":"llm","lobe":"frontal","color":"#6784b0","indegree":3,"alsoIn":[],"x":0.8948,"y":0.5632,"text":"The paper tackles the 'adapter caching problem' in distributed LLM serving: given an anticipated workload of many heterogeneous LoRA adapters, place them across GPUs and set per-GPU A_max so the workload is served with the minimum number of GPUs while avoiding request starvation and GPU memory errors. It proposes a data-driven pipeline of three components \u2014 a Digital Twin that emulates LLM-adapter serving (continuous batching, KV-cache allocation, adapter swap), distilled ML surrogate models trained on DT-generated data to predict per-GPU throughput and starvation risk, and a greedy First-Fit-Decreasing placement algorithm that uses those predictions to drive GPUs to their optimal packing point Max_pack. The Digital Twin reproduces real behavior with under ~5% throughput SMAPE while running up to 90\u00d7 faster than real benchmarking, and the full pipeline reduces the number of GPUs needed versus throughput-agnostic baselines and dLoRA while remaining feasible (no starvation/memory errors)."},{"id":"arxiv:2503.02143","kind":"paper","n":120,"label":"Four Principles for Physically Interpretable World Models","authors":"","year":null,"url":"https://arxiv.org/abs/2503.02143","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":1,"alsoIn":[],"x":0.2496,"y":0.6715,"text":"This position/principles paper argues for shifting from physically *informed* to physically *interpretable* world models \u2014 models whose latent space maps to meaningful physical variables and whose latent dynamics emulate real physical processes. It crystallizes four design principles leveraging symbolic knowledge: (1) functionally organizing the latent space by physical intent, (2) learning aligned invariant/equivariant representations, (3) integrating multiple forms/strengths of supervision, and (4) partitioning generative outputs into separately verifiable segments. Each principle is given a formal definition/loss and lightweight validation on the Cart Pole and Lunar Lander OpenAI Gym benchmarks using a VAE+LSTM world model, measuring state-prediction MSE, reconstruction SSIM, and model size."},{"id":"arxiv:2103.03725","kind":"paper","n":121,"label":"Predictive Coding Can Do Exact Backpropagation on Convolutional and Recurrent Neural Networks","authors":"Tommaso Salvatori, Yuhang Song, Thomas Lukasiewicz, Rafal Bogacz, Zhenghua Xu","year":2021,"url":"https://arxiv.org/abs/2103.03725","region":"t20","hemi":"llm","lobe":"frontal","color":"#67b0b0","indegree":6,"alsoIn":[],"x":0.6316,"y":0.6503,"text":"The paper addresses whether predictive coding networks (PCNs), a biologically plausible model of cortical processing, can exactly replicate backpropagation (BP) on complex architectures rather than only fully connected networks. Building on the Z-IL (zero-divergence inference learning) variant of inference learning, the authors define convolutional and many-to-one recurrent versions of PCNs and mathematically prove that Z-IL produces weight updates strictly equivalent to BP on CNNs and RNNs (Theorems 1-4). They also analyze running time, showing Z-IL needs only l_max inference steps versus the ~100-200 steps required by standard IL, making it orders of magnitude faster than IL and only slightly slower than BP. This is presented as the first biologically plausible algorithm to exactly match BP's accuracy on such architectures."},{"id":"arxiv:2510.12966","kind":"paper","n":122,"label":"3-Model Speculative Decoding","authors":"","year":null,"url":"https://arxiv.org/abs/2510.12966","region":"t9","hemi":"llm","lobe":"parietal","color":"#97b067","indegree":1,"alsoIn":[],"x":0.8709,"y":0.8291,"text":"Speculative decoding (SD) speeds up LLM inference but is bottlenecked by a draft-size vs. acceptance-rate trade-off: smaller drafts are faster but diverge more from the target, lowering acceptance. The paper introduces Pyramid Speculative Decoding (PyramidSD), a training-free three-model hierarchy that inserts an intermediate 'qualifier' model between draft and target, using fuzzy (divergence-threshold) acceptance at each of two speculative stages to bridge the distributional gap. On CommonsenseQA with the Llama 3.2/3.1 family (1B draft / 3B qualifier / 8B target) on an RTX 4090, the fuzzy variant (PSDF) reaches up to 1.91\u00d7 speedup over standard SD (124.13 tok/s), while the assisted variant (PSDA) gives up to 1.44\u00d7 with low variance and minimal quality loss."},{"id":"arxiv:2604.16592","kind":"paper","n":123,"label":"Human Cognition in Machines: A Unified Perspective of World Models","authors":"","year":null,"url":"https://arxiv.org/abs/2604.16592","region":"t7","hemi":"llm","lobe":"occipital","color":"#adb067","indegree":2,"alsoIn":[],"x":0.6794,"y":0.5264,"text":"This survey reframes contemporary world models through Cognitive Architecture Theory (CAT), classifying prior work by the seven cognitive functions it innovates on\u2014memory, perception, language, reasoning, imagination, motivation, and metacognition\u2014rather than by application domain. Applying this lens across video, embodied, and a newly proposed 'epistemic' world-model category, the authors find that motivation (especially intrinsic motivation) and metacognition are drastically under-researched in current systems. They propose a unified world-model framework as a conceptual roadmap and argue that active inference (for motivation) and language-based global workspaces in agentic discovery systems (for metacognition) offer concrete paths to fill these gaps."},{"id":"arxiv:2605.02900","kind":"paper","n":124,"label":"Safety in Embodied AI: A Survey of Risks, Attacks, and Defenses","authors":"","year":null,"url":"https://arxiv.org/abs/2605.02900","region":"t7","hemi":"llm","lobe":"occipital","color":"#adb067","indegree":0,"alsoIn":[],"x":0.6959,"y":0.5819,"text":"This survey provides a comprehensive, structured review of safety research in embodied AI, synthesizing insights from over 500 papers on attacks and defenses across the full embodied pipeline: perception, cognition, planning, action & interaction, and agentic systems. It introduces a multi-level taxonomy organized around a 'capability\u2013risk duality' principle, where each capability layer expands the attack surface, and inner-layer vulnerabilities cascade outward. The work catalogs adversarial, backdoor, jailbreak, sensor/hardware, tool-misuse, memory-poisoning, and cascading-failure threats alongside corresponding defenses (robust training, robust inference, anti-spoofing, runtime guardrails). It identifies overlooked challenges including the fragility of multimodal fusion, instability of planning under jailbreak, and untrustworthy human\u2013agent interaction, and lays out open problems and future trends."},{"id":"arxiv:2505.21996","kind":"paper","n":125,"label":"Learning World Models for Interactive Video Generation","authors":"Taiye Chen, et al.","year":2025,"url":"https://arxiv.org/abs/2505.21996","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":4,"alsoIn":[],"x":0.3248,"y":0.5673,"text":"The paper tackles long-horizon interactive video generation as world modeling, identifying two coupled failure modes in autoregressive video models: irreducible compounding errors and insufficient memory mechanisms that break spatiotemporal consistency. The authors augment an image-to-video latent diffusion model with action conditioning (AdaLN) and Diffusion Forcing for autoregressive rollout, then propose VRAG \u2014 Video Retrieval Augmented Generation \u2014 which retrieves relevant historical frames via global-state (coordinate/pose) similarity and adds explicit global state conditioning. On a self-collected Minecraft benchmark, VRAG achieves the best world-coherence (SSIM 0.506) and lowest compounding error (SSIM 0.349), outperforming long-context (YaRN), history-buffer, neural-memory (Infini-attention), and Frame Pack baselines, and the approach generalizes to RealEstate10K. A key finding is that LLM long-context/RAG techniques transfer poorly to video because current video diffusion models have weak in-context learning."},{"id":"arxiv:2507.05201","kind":"paper","n":126,"label":"MedGemma Technical Report","authors":"Google Research and Google DeepMind","year":2025,"url":"https://arxiv.org/abs/2507.05201","region":"t13","hemi":"llm","lobe":"parietal","color":"#6bb067","indegree":2,"alsoIn":[],"x":0.7632,"y":0.179,"text":"MedGemma is a collection of open medical vision-language foundation models built on Gemma 3 (4B multimodal and 27B text-only, plus a 27B multimodal variant), trained via vision-encoder enhancement, multimodal pretraining, and post-training (distillation + RL) on medical image and text data. The models substantially outperform same-sized Gemma 3 baselines across medical text QA, image classification, VQA, CXR report generation, and agentic tasks while retaining general capabilities. Fine-tuning further boosts performance in subdomains, and the paper also introduces MedSigLIP, a 400M medical image encoder derived from SigLIP that matches or exceeds specialized medical encoders. For OOD tasks, MedGemma gains 2.6-10% on medical multimodal QA, 15.5-18.1% on CXR finding classification, and 10.8% on agentic evaluations versus base models."},{"id":"arxiv:2202.08906","kind":"paper","n":127,"label":"ST-MoE: Designing Stable and Transferable Sparse Expert Models","authors":"Barret Zoph et al.","year":2022,"url":"https://arxiv.org/abs/2202.08906","region":"t16","hemi":"llm","lobe":"frontal","color":"#67b084","indegree":3,"alsoIn":[],"x":0.7378,"y":0.4525,"text":"The paper studies why sparse Mixture-of-Experts (MoE) language models suffer training instabilities and uncertain fine-tuning quality, and offers a design guide to fix both. It introduces the router z-loss, which penalizes large router logits to stabilize training without hurting (and slightly improving) quality, and presents fine-tuning, capacity-factor, and routing recommendations along with a token-routing analysis. Using these techniques, the authors train ST-MoE-32B, a 269B-parameter sparse model FLOP-matched to a 32B dense Transformer. It becomes the first sparse model to reach state-of-the-art transfer-learning performance across reasoning, summarization, closed-book QA, and adversarial NLP benchmarks (e.g., SuperGLUE test score 91.2)."},{"id":"arxiv:2308.12284","kind":"paper","n":128,"label":"D4: Improving LLM Pretraining via Document De-Duplication and Diversification","authors":"Tirumala et al.","year":2023,"url":"https://arxiv.org/abs/2308.12284","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":2,"alsoIn":[],"x":0.7787,"y":0.6799,"text":"Addresses LLM pretraining data selection beyond simple MinHash de-duplication, where data is normally randomly sampled in a single pass. The authors introduce D4, which applies semantic de-duplication (SemDeDup) then re-clusters and prunes prototypical points (SSL Prototypes) in a pre-trained 125M OPT embedding space to diversify the training set. On OPT models up to 6.7B trained on 100B tokens, D4 yields ~20% training-efficiency gains and up to a 2% average 0-shot downstream accuracy improvement across 16 NLP tasks. They also show that intelligently repeating D4-selected data over multiple epochs beats one-pass training on randomly selected new tokens, challenging the standard single-epoch practice."},{"id":"arxiv:2208.01769","kind":"paper","n":129,"label":"Deep Reinforcement Learning for Multi-Agent Interaction","authors":"","year":null,"url":"https://arxiv.org/abs/2208.01769","region":"t22","hemi":"wm","lobe":"temporal","color":"#4f849a","indegree":10,"alsoIn":[],"x":0.1335,"y":0.3063,"text":"This is an overview article from the University of Edinburgh's Autonomous Agents Research Group describing its research portfolio in deep reinforcement learning (RL) and multi-agent reinforcement learning (MARL). It surveys the group's contributions across five strands: scalable MARL with coordinated policies and emergent communication, ad hoc teamwork and agent modelling, sample-efficient single-agent RL, autonomous driving with interpretable goal recognition/planning, and quantum-secure authentication via multi-agent interaction. It catalogs named methods (SEAC, SePS, GPL, LIAM, DeRL, IGP2, GOFI, GRIT, AMI) and open-source repositories/environments (EPyMARL, LBF, RWARE, PressurePlate, MIDGARD), then outlines open problems in RL generalisation, causal RL, ad hoc teamwork, and driving."},{"id":"arxiv:2512.02419","kind":"paper","n":130,"label":"The brain-AI convergence: Predictive and generative world models for general-purpose computation","authors":"(authors listed on arXiv)","year":2025,"url":"https://arxiv.org/abs/2512.02419","region":"t5","hemi":"llm","lobe":"parietal","color":"#b09e67","indegree":8,"alsoIn":[],"x":0.6615,"y":0.5512,"text":"This Perspective provides a cross-domain computational comparison between brain circuits (neocortex and cerebellum) and modern AI (transformers, LLMs) through the lens of world-model-based computation. The authors identify shared mechanisms: both brain regions and AI predict future world events from past inputs, construct internal world models via prediction-error learning, and repurpose these models for understanding (sensory processing) and generation (motor processing). The paper concludes that convergent evolution between brain and AI rests on three principles\u2014prediction-error learning, single-circuit computation of prediction/understanding/generation, and adaptability via attention mechanisms and mixture-of-experts\u2014constituting a core computational foundation for diverse functions from uniform circuit architectures."},{"id":"arxiv:2603.28489","kind":"paper","n":131,"label":"Video Generation Models as World Models: Efficient Paradigms, Architectures and Algorithms","authors":"(authors listed on arXiv)","year":2026,"url":"https://arxiv.org/abs/2603.28489","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":10,"alsoIn":[],"x":0.4029,"y":0.7054,"text":"A survey positioning video generation models as world simulators and arguing that computational efficiency is a prerequisite for turning them into real-time, long-horizon world models. It introduces a three-dimensional taxonomy \u2014 efficient modeling paradigms (diffusion distillation, autoregressive/hybrid AR-diffusion, streaming causal diffusion), efficient architectures (hierarchical/VAE compression, long-context memory, sparse/linear/windowed attention, RoPE extrapolation), and efficient inference (parallelism, caching, pruning, quantization). It then maps these efficiency techniques onto interactive applications (autonomous driving, embodied AI, game/world simulation) across data synthesis, interactive simulation, and generative planning, and identifies open frontiers in physics-aware, long-horizon consistency."},{"id":"arxiv:2511.12239","kind":"paper","n":132,"label":"Beyond World Models: Rethinking Understanding in AI Models","authors":"Tarun Gupta, Danish Pruthi","year":2025,"url":"https://arxiv.org/abs/2511.12239","region":"t6","hemi":"llm","lobe":"temporal","color":"#b0a967","indegree":4,"alsoIn":[],"x":0.7019,"y":0.6542,"text":"This position/theory paper critically examines whether the 'world model' framework \u2014 internal representations that track objects, states, and state-transition rules \u2014 adequately characterizes human-level understanding in AI models. Using three case studies from philosophy of science (Hofstadter's domino-chain prime-testing computer, the distinction between verifying vs. understanding mathematical proofs per Poincar\u00e9/Avigad, and Popper's account of understanding Bohr's atomic theory via its 'problem situation'), the authors argue that even a perfect world model that tracks physical/logical states fails to capture the abstract concepts (e.g., primality), the motivating insights, and the explanatory problem-context that constitute genuine understanding. They also rebut the counterargument that abstract concepts could be encoded as states, arguing this destroys the framework's falsifiability and explanatory power. The thesis is not that AI cannot understand, but that the world-model conception is an inadequate theoretical lens for understanding."},{"id":"arxiv:2502.03373","kind":"paper","n":133,"label":"Demystifying Long Chain-of-Thought Reasoning in LLMs","authors":"Edward Yeo et al.","year":2025,"url":"https://arxiv.org/abs/2502.03373","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":9,"alsoIn":[],"x":0.8089,"y":0.8151,"text":"The paper systematically investigates how long chain-of-thought (CoT) reasoning emerges in LLMs through supervised fine-tuning (SFT) and reinforcement learning (RL), using Llama-3.1-8B and Qwen2.5-Math-7B trained on the MATH dataset. It shows that SFT on emergent long-CoT data (distilled from QwQ-32B-Preview) scales to a higher accuracy ceiling than short CoT and provides a better RL starting point, but that RL does not stably scale CoT length without reward shaping; the authors introduce a cosine length-scaling reward plus an n-gram repetition penalty to stabilize growth. They further show noisy web-extracted (WebInstruct) data with rule-based filtering boosts out-of-distribution/STEM performance, and that core skills like self-validation already exist in base models so RL mostly recombines pretrained abilities rather than creating new 'aha moments'."},{"id":"arxiv:1011.0063","kind":"paper","n":134,"label":"Multiscales and cascade in isotropic turbulence","authors":"","year":null,"url":"https://arxiv.org/abs/1011.0063","region":"t36","hemi":"llm","lobe":"frontal","color":"#b06793","indegree":4,"alsoIn":[],"x":0.6585,"y":0.5204,"text":"The paper tackles the energy cascade problem in fully developed isotropic turbulence, arguing that Richardson's cascade and Kolmogorov's picture have never been derived deductively. Using an 'explicit map method' applied to a previously derived self-preserving turbulence scaling equation, the author reduces the hierarchy of eddy length scales to a recursion relation that is shown to be equivalent to the logistic map, with a control parameter that depends only on a relative Reynolds number. This yields a Feigenbaum-type bifurcation diagram in which the Richardson cascade appears as an infinite sequence of period-doubling bifurcations leading to chaos. The work claims to give the first visual/deductive evidence of the multiscale Richardson cascade and links the laminar-to-turbulent transition to the period-doubling route to chaos."},{"id":"arxiv:2605.17288","kind":"paper","n":135,"label":"When Efficiency Backfires: Cascading LLMs Trigger Cascade Failure under Adversarial Attack","authors":"","year":null,"url":"https://arxiv.org/abs/2605.17288","region":"t8","hemi":"llm","lobe":"frontal","color":"#a2b067","indegree":3,"alsoIn":[],"x":0.8544,"y":0.6515,"text":"LLM cascade systems (e.g., FrugalGPT) reduce inference cost by routing queries through lightweight models before escalating to stronger ones, but this sequential architecture introduces structural vulnerabilities absent in standalone models. The authors propose a joint-target adversarial attack framework that decomposes adversarial suffixes into components targeting both prediction models and decision modules, with coordinated alternating optimization and adaptive pass-rate control to simultaneously degrade accuracy and inflate cost. Experiments across 9 datasets, 12 LLM architectures, and 33 cascade configurations show attacks can reduce accuracy by up to 84.6%, inflate token cost by 148.9%, increase execution time by 108.9%, and boost jailbreak success rates by up to 81.4%, with attacks remaining effective under black-box transfer and only partially mitigated by existing defenses."},{"id":"arxiv:1712.04602","kind":"paper","n":136,"label":"On the organization of grid and place cells: Neural de-noising via subspace learning","authors":"","year":null,"url":"https://arxiv.org/abs/1712.04602","region":"t28","hemi":"llm","lobe":"frontal","color":"#7667b0","indegree":4,"alsoIn":[],"x":0.5927,"y":0.3349,"text":"The paper analyzes the joint neural code for spatial location formed by grid cells (MEC) and place cells (hippocampus) through a coding-theory lens, treating their combined population activity as a 'hybrid code' whose codewords are generated by stimulating the network at discrete locations. The authors apply a biologically plausible, modified Oja subspace-learning rule and high-capacity associative-memory de-noising networks (clustered by grid module vs. un-clustered/randomly-clustered) to correct noise-corrupted population activity, and characterize how code construction parameters (phase/orientation redundancy, uniform vs. non-uniform allocation of grid cells to modules) affect normalized rank, minimum distance, and rate. Simulations show that de-noising networks whose constraint neurons are clustered by grid module dramatically improve fidelity of the spatial representation, but only when grid cells have sufficient phase redundancy. The learned connectivity further predicts that effective place-cell-to-grid-module connectivity decreases for larger place fields, implying inter-hippocampal-MEC connectivity decreases along the dorsoventral axis."},{"id":"arxiv:2504.09590","kind":"paper","n":137,"label":"Efficient LLM Serving on Hybrid Real-time and Best-effort Requests","authors":"","year":null,"url":"https://arxiv.org/abs/2504.09590","region":"t24","hemi":"llm","lobe":"frontal","color":"#6784b0","indegree":5,"alsoIn":[],"x":0.9247,"y":0.6085,"text":"Production LLM services concurrently handle latency-critical real-time (RT) requests and throughput-oriented best-effort (BE) requests, but current systems dedicate separate GPU clusters to each, causing poor utilization during off-peak RT periods. BROS collocates RT/BE requests on shared GPUs using a dynamic priority-based packing algorithm that schedules at the iteration level to meet RT token-latency SLOs (TTFT and TPOT) while maximizing BE throughput, plus a bidirectional KV cache layout that lets RT and BE requests share memory blocks to avoid costly recomputation or swapping. Experiments on OPT-13B/30B and Llama-3-70B show BROS reduces RT latency by up to 74.20% versus vLLM with only 11.29% BE throughput loss, and improves TTFT SLO attainment by up to 36.38\u00d7."},{"id":"arxiv:2605.29930","kind":"paper","n":138,"label":"Toward AI That Understands Self and Others: A World-Model Theory of Cognitive Diversity and Alignment","authors":"(authors listed on arXiv)","year":2026,"url":"https://arxiv.org/abs/2605.29930","region":"t6","hemi":"llm","lobe":"temporal","color":"#b0a967","indegree":1,"alsoIn":[],"x":0.6764,"y":0.6082,"text":"This paper addresses why different subjects reach different conclusions from the same observation sequence, arguing that observation is not yet inference and that a target becomes inferentially admissible only when an approximate sufficient statistic can be constructed for it under finite constraints. It formalizes this as the Multi-Phase Inference Assumption (MIA) and Multi-Phase Inference Mechanism (MIM), introducing the phase-formation space, foregrounding gradient field, and alignment map \u03a6 to describe how heterogeneous world models can be made mutually processable without collapsing into a single representation. The framework redefines AI alignment as processability rather than agreement, positioning cognitive diversity as distributed error detection and current methods like RLHF/DPO/Constitutional AI as restricted to the value layer of a three-layer alignment hierarchy. The paper is a foundation paper with formal definitions and five verifiable hypotheses but no experimental validation."},{"id":"arxiv:2412.03568","kind":"paper","n":139,"label":"The Matrix: Infinite-Horizon World Generation with Real-Time Moving Control","authors":"Feng, Zhang et al.","year":2024,"url":"https://arxiv.org/abs/2412.03568","region":"t15","hemi":"wm","lobe":"occipital","color":"#4f9a62","indegree":1,"alsoIn":[],"x":0.3818,"y":0.6392,"text":"The paper presents The Matrix, a 2.7B-parameter foundational world simulator that generates infinitely long 720p (1280\u00d7720) real-scene video streams with real-time, frame-level interactive control in first- and third-person perspectives. It is fine-tuned from a pre-trained 2.3B video DiT backbone using a small amount of supervised AAA game data (Forza Horizon 5, Cyberpunk 2077) collected via a GameData platform plus large-scale unsupervised real-world footage. Three novel components \u2014 an Interactive Module for keyboard-to-text frame-level control, a Shift-Window Denoising Process Model (Swin-DPM) for infinite-length streaming generation, and a Stream Consistency Model (SCM) for real-time acceleration \u2014 let it run at 8\u201316 FPS with zero-shot generalization to out-of-distribution scenes (e.g., driving a BMW X3 indoors)."},{"id":"arxiv:2206.04114","kind":"paper","n":140,"label":"Deep Hierarchical Planning from Pixels (Director)","authors":"Hafner et al.","year":2022,"url":"https://arxiv.org/abs/2206.04114","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":1,"alsoIn":[],"x":0.1667,"y":0.422,"text":"Director tackles long-horizon, sparse-reward control where flat RL agents fail by learning hierarchical behavior entirely from pixels via planning inside the latent space of a learned world model (PlaNet/DreamerV2 RSSM). A manager policy picks latent goals every K=8 steps in a compact discrete code space produced by a goal autoencoder, while a goal-conditioned worker learns to reach those goals using a max-cosine similarity reward; the manager additionally maximizes task reward plus an autoencoder-reconstruction-error exploration bonus for temporally-abstract exploration. Director solves four egocentric Ant Maze navigation tasks and Visual Pin Pad tasks where Dreamer and Plan2Explore baselines fail, and generalizes across Atari, DMLab, Crafter, and Control Suite. Because latent goals decode back into images, the agent's subgoal decisions are human-interpretable."},{"id":"arxiv:2411.17470","kind":"paper","n":141,"label":"Towards Precise Scaling Laws for Video Diffusion Transformers","authors":"Yin et al.","year":2024,"url":"https://arxiv.org/abs/2411.17470","region":"t32","hemi":"wm","lobe":"frontal","color":"#8b4f9a","indegree":1,"alsoIn":[],"x":0.3694,"y":0.2399,"text":"The paper investigates whether scaling laws exist for video diffusion transformers (video DiT) and how to derive them precisely, motivated by the high cost of training large video generators. The authors confirm scaling laws hold, but find video DiT is far more sensitive to learning rate and batch size than language models, so they derive theory-driven power-law equations predicting the optimal batch size and learning rate as functions of model size N and training tokens T, plus a generalized validation-loss law L(T,N). Using these optimal hyperparameters, they predict optimal model size under a compute budget and validate extrapolations on 0.72B and 1.07B models trained with the Cross-DiT architecture on Panda-70M. Optimal-hyperparameter fitting yields more parameter-efficient models (39.9% fewer parameters / 40.1% lower inference cost at comparable loss) and substantially more accurate loss prediction than fixed suboptimal settings."},{"id":"arxiv:2411.07690","kind":"paper","n":142,"label":"World Models: The Safety Perspective","authors":"","year":null,"url":"https://arxiv.org/abs/2411.07690","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":1,"alsoIn":[],"x":0.3274,"y":0.8158,"text":"This survey reviews the state of the art in World Models (WM) for embodied AI agents through the lens of safety and trustworthiness. It builds a chronological taxonomy of WM techniques across four categories \u2014 RNN-based, transformer-based, diffusion-based, and other methods (e.g., JEPA) \u2014 and surveys their evaluation metrics. Through concrete failure cases in autonomous driving (vehicles spawned off drivable areas in TrafficGen, traffic-rule violations in CTG, implausible road markings and vehicles disappearing between frames in DriveDreamer, and a MILE-agent collision at t=3.2s in CARLA), it argues current WMs do not meet minimum safety requirements, and proposes a research agenda for trustworthy WMs."},{"id":"arxiv:1907.05600","kind":"paper","n":143,"label":"Generative Modeling by Estimating Gradients of the Data Distribution","authors":"Yang Song, Stefano Ermon","year":2019,"url":"https://arxiv.org/abs/1907.05600","region":"t3","hemi":"wm","lobe":"occipital","color":"#9a714f","indegree":2,"alsoIn":[],"x":0.387,"y":0.1262,"text":"The paper addresses generative modeling by learning the score (gradient of log-density) of the data distribution and sampling via Langevin dynamics, avoiding adversarial training and MCMC during training. It identifies two obstacles\u2014the manifold hypothesis making scores ill-defined and inaccurate score estimation plus slow Langevin mixing in low-density regions\u2014and solves them by perturbing data with multiple Gaussian noise levels, training a single Noise Conditional Score Network (NCSN) via denoising score matching to estimate scores at all levels, then sampling with annealed Langevin dynamics that anneals noise from high to low. On CIFAR-10 the method reaches a new state-of-the-art unconditional inception score of 8.87 and a competitive FID of 25.32, with comparable-quality samples on MNIST and CelebA and demonstrated image inpainting."},{"id":"arxiv:2510.01539","kind":"paper","n":144,"label":"Executable Counterfactuals: Improving LLMs' Causal Reasoning Through Code","authors":"Anonymous et al.","year":2025,"url":"https://arxiv.org/abs/2510.01539","region":"t6","hemi":"llm","lobe":"temporal","color":"#b0a967","indegree":1,"alsoIn":[],"x":0.7019,"y":0.7158,"text":"The paper argues that existing evaluations of LLM counterfactual reasoning skip the abduction step, collapsing counterfactual (Pearl Level 3) into interventional (Level 2) reasoning and overstating model ability. It introduces 'executable counterfactuals,' a framework that operationalizes the full abduction-intervention-prediction rollout via code (Python functions with hidden latent variables) and GSM-style math word problems, enabling scalable synthetic data generation with controllable OOD structure. SOTA models drop 25-40% from interventional to counterfactual reasoning; SFT distillation improves in-distribution but degrades out-of-distribution, whereas RLVR (GRPO) induces the core cognitive behaviors and generalizes across code structures and to natural-language math, roughly 1.5-2x over the base model."},{"id":"arxiv:2601.21998","kind":"paper","n":145,"label":"","authors":"","year":null,"url":"https://arxiv.org/abs/2601.21998","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":8,"alsoIn":[],"x":0.2458,"y":0.7766,"text":"The paper tackles representation entanglement in feedforward VLA policies, which jointly compress visual understanding, dynamics, and motor control into one supervision signal, hurting sample efficiency and generalization. It introduces LingBot-VA, an autoregressive diffusion world model that interleaves video and action tokens in a single causal sequence via a dual-stream Mixture-of-Transformers (video stream from Wan2.2-5B, smaller action stream), predicting future visual states by flow matching and decoding actions by inverse dynamics, with KV-cache memory, partial-denoising via Noisy History Augmentation, and FDM-grounded asynchronous inference. On RoboTwin 2.0 it reaches 92.93% (Easy)/91.55% (Hard) average over 50 tasks and 98.5% average on LIBERO, beating \u03c00, \u03c00.5, X-VLA, and Motus. It also shows strong sample efficiency (15.6% higher progress on Make Breakfast at 10 demos), long-horizon temporal memory, and generalization, using only ~50 demos for real-world adaptation."},{"id":"arxiv:2305.17198","kind":"paper","n":146,"label":"A Model-Based Solution to the Offline Multi-Agent Reinforcement Learning Coordination Problem","authors":"Paul Barde, Jakob Foerster, Derek Nowrouzezahrai, Amy Zhang","year":2024,"url":"https://arxiv.org/abs/2305.17198","region":"t22","hemi":"llm","lobe":"temporal","color":"#4f849a","indegree":3,"alsoIn":[],"x":0.642,"y":0.8205,"text":"The paper identifies and formalizes the 'offline coordination problem' in multi-agent reinforcement learning, decomposing it into two challenges: strategy agreement (SA, agents collectively selecting one of several equivalent optimal team strategies) and strategy fine-tuning (SFT, calibrating individual behaviors to one another). It shows that prevalent model-free offline MARL methods fail at these challenges even in toy environments, and argues the root cause is the absence of inter-agent interactions during offline learning. It proposes MOMA-PPO, the first model-based offline MARL method, which learns a centralized world-model ensemble and uses Dyna-like synthetic rollouts to train MAPPO agents, letting them interact through the world model. MOMA-PPO solves the Iterated Coordination Game and significantly outperforms model-free baselines on Multi-Agent MuJoCo Reacher and Ant tasks, even under severe partial observability and with learned world models."},{"id":"arxiv:2404.12377","kind":"paper","n":147,"label":"RoboDreamer: Learning Compositional World Models for Robot Imagination","authors":"Zhou, Du, Chen, Li, Yeung, Gan","year":2024,"url":"https://arxiv.org/abs/2404.12377","region":"t0","hemi":"wm","lobe":"frontal","color":"#9a4f4f","indegree":1,"alsoIn":[],"x":0.2937,"y":0.7872,"text":"Text-to-video world models for robotics fail to generalize to unseen language instructions, especially novel object-action-spatial combinations not seen during training. RoboDreamer learns a compositional world model by using a constituency parser to factorize each instruction into verb/action phrases and prepositional/relation phrases, then conditions a product of diffusion score functions on each parsed primitive so that novel instructions generalize as long as each component is in-distribution. The same factorization lets the model compose extra multimodal conditions (goal images, goal sketches) at inference time. On RT-1 video generation it substantially beats monolithic baselines on human-rated unseen tasks (81.3 vs 46.9 for AVDC), and on RLBench it achieves the best average manipulation success rate (49.3%)."},{"id":"arxiv:2210.07729","kind":"paper","n":148,"label":"Model-Based Imitation Learning for Urban Driving","authors":"Hu et al. (Wayve)","year":2022,"url":"https://arxiv.org/abs/2210.07729","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":1,"alsoIn":[],"x":0.3067,"y":0.6114,"text":"MILE addresses motion planning for urban autonomous driving by jointly learning a world model and a driving policy purely from an offline corpus of expert demonstrations, with no online environment interaction or access to a reward. The method uses 3D geometry as an inductive bias by lifting image features to a bird's-eye-view representation and compressing them into a compact 512-dim latent space, then models temporal dynamics via a variational recurrent latent-state model that predicts actions, BeV segmentation, and future states. On the CARLA simulator in an unseen town and weather, MILE achieves a driving score of 61.1, a 31% relative improvement over the prior state-of-the-art LAV (46.5), using only a single front RGB camera. It can also predict diverse, plausible futures and execute complex maneuvers (e.g., roundabouts) from plans entirely imagined in latent space, sustaining performance with up to 30% of driving done in imagination."},{"id":"arxiv:2209.00588","kind":"paper","n":149,"label":"Transformers are Sample-Efficient World Models (IRIS)","authors":"Micheli, Alonso & Fleuret","year":2023,"url":"https://arxiv.org/abs/2209.00588","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":5,"alsoIn":[],"x":0.1771,"y":0.4707,"text":"Deep RL agents are notoriously sample inefficient, which limits real-world deployment; learning a policy purely inside a world model frees it from interaction constraints but demands an accurate world model. The paper introduces IRIS, a data-efficient agent whose world model is composed of a discrete autoencoder (VQVAE-style) that turns frames into image tokens and a GPT-like autoregressive Transformer that models environment dynamics over those tokens, with the policy trained entirely in imagination via a DreamerV2-style actor-critic. On the Atari 100k benchmark (\u22482 hours of gameplay, 100k actions), IRIS reaches a mean human-normalized score of 1.046 and beats humans on 10 of 26 games, a new state of the art for methods without lookahead search. The world model learns game mechanics well enough to produce pixel-perfect rollouts in some games and the approach scales to 10M steps for a mean score of 7.488."},{"id":"arxiv:2305.18264","kind":"paper","n":150,"label":"Gen-L-Video: Multi-Text to Long Video Generation via Temporal Co-Denoising","authors":"Fu-Yun Wang et al.","year":2023,"url":"https://arxiv.org/abs/2305.18264","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":4,"alsoIn":[],"x":0.4025,"y":0.3828,"text":"Text-driven video diffusion models are typically confined to very short clips (<24 frames) under a single text condition. This paper introduces Gen-L-Video, a training-free paradigm that treats a long video as a collection of temporally overlapping short clips and approximates the long-video denoising trajectory by jointly denoising these overlapping clips (Temporal Co-Denoising), merging them via a closed-form weighted average per frame. It extends three mainstream paradigms (pretrained t2v, tuning-free t2v, one-shot-tuning t2v) to generate/edit videos of hundreds of frames conditioned on multiple text prompts while preserving content consistency, and adds a Bi-Directional Cross-Frame Attention and clip identifiers to improve consistency. Experiments on a 66-video benchmark show Gen-L-Video improves frame consistency (93.18 vs 91.65) and is preferred by humans 85.38% vs 14.62% over isolated denoising."},{"id":"arxiv:2605.23972","kind":"paper","n":151,"label":"Why We Need World Models for AGI: Where LLMs Fail and How World Models May Outperform","authors":"Anonymous","year":2026,"url":"https://arxiv.org/abs/2605.23972","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":6,"alsoIn":[],"x":0.645,"y":0.8732,"text":"The paper argues that LLM limitations in causal reasoning, persistent state tracking, and long-horizon planning stem from an objective-level mismatch between next-token sequence prediction (over observation space X) and reasoning over latent environment dynamics (state space S). It introduces Latent Dynamics Inference (LDI), a conceptual framing of language/multimodal data as lossy partial observations of latent state trajectories, and FLUX, a two-player sequential game specified in natural language that is compiled into an explicit state-transition simulator. In a controlled case study, a tabular Q-learning agent operating in the extracted latent state space wins ~79% of games versus ~11% for frontier LLMs (GPT-4o, DeepSeek-R1, Gemini 1.5 Pro), with ~10% of LLM losses due to invalid-move forfeits. Qualitative analysis attributes LLM failures to sum blindness, row-length miscounting, and horizon myopia\u2014consistent with the absence of a persistently updated latent state."},{"id":"arxiv:2407.15549","kind":"paper","n":152,"label":"Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMs","authors":"","year":null,"url":"https://arxiv.org/abs/2407.15549","region":"t8","hemi":"llm","lobe":"frontal","color":"#a2b067","indegree":6,"alsoIn":[],"x":0.838,"y":0.6623,"text":"LLMs retain harmful capabilities that fine-tuning suppresses but fails to remove, leaving them vulnerable to jailbreaks, backdoors, and knowledge re-extraction. The paper introduces targeted latent adversarial training (LAT), in which an adversary perturbs the model's residual-stream activations to actively elicit a specific undesirable behavior, and the model is then fine-tuned to resist those perturbations. Applied on top of refusal training, DPO, WHP, gradient ascent, and RMU, targeted LAT improves robustness to jailbreaks (beating R2D2 with orders of magnitude less compute), removes RLHF backdoors without knowing the trigger, and makes unlearning more robust to re-learning. Across all three settings it reduces harmful behavior with little to no loss of general capability (e.g., <1 point MMLU drop in backdoor removal)."},{"id":"frontiersin.org:d1ec3c5c2c89","kind":"paper","n":153,"label":"Map-like representations of an abstract conceptual space in the human brain (audiovisual grid code)","authors":"Simone Vigano, Manuela Piazza","year":2020,"url":"https://www.frontiersin.org/journals/neural-circuits/articles/10.3389/fncir.2022.924016/full","region":"t28","hemi":"llm","lobe":"frontal","color":"#7667b0","indegree":6,"alsoIn":[],"x":0.5568,"y":0.3531,"text":"This hypothesis-and-theory article argues that grid-like representations may be a universal computational component of all perception and cognition, extending beyond the classical hippocampal-entorhinal system to sensory cortices (S1, V2) and frontal areas (OFC, vmPFC, ACC). The authors synthesize evidence from rodent electrophysiology, human fMRI, and computational modeling (RNNs, successor representations) to propose that grid cells serve generalized spatial localization across physical and abstract conceptual spaces. They outline four experimentally testable predictions: grid cells in auditory cortex, human somatosensory/visual cortices, frontal cortex, and universal egocentric-to-allocentric transformation functions across sensory cortices."},{"id":"arxiv:2506.09171","kind":"paper","n":154,"label":"Improving LLM Agent Planning with In-Context Learning via Atomic Fact Augmentation and Lookahead Search","authors":"Samuel Holt, Max Ruiz Luyten, Thomas Pouplin, Mihaela van der Schaar","year":2025,"url":"https://arxiv.org/abs/2506.09171","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":4,"alsoIn":[],"x":0.6316,"y":0.9267,"text":"LLM agents struggle to plan effectively in partially observable, long-horizon environments when search is unguided or recent history is insufficient. LWM-Planner extracts task-critical atomic facts from episodic trajectories, validates them with a predictive-consistency filter, and uses the resulting fact set to condition LLM-driven action proposal, single-step latent world-model simulation, and state-value estimation within a recursive depth-limited lookahead search\u2014all without parameter updates. Across TextFrozenLake, CrafterMini, and ALFWorld, LWM-Planner consistently improves cumulative return over ReAct, Reflexion, ToT, and RAP baselines, with ablations showing that fact-grounded lookahead is the dominant contributor to gains while unguided search can even harm performance."},{"id":"arxiv:1912.01603","kind":"paper","n":155,"label":"Dream to Control: Learning Behaviors by Latent Imagination (Dreamer)","authors":"Hafner et al.","year":2020,"url":"https://arxiv.org/abs/1912.01603","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":3,"alsoIn":[],"x":0.1408,"y":0.396,"text":"Dreamer is a model-based reinforcement learning agent that learns long-horizon continuous-control behaviors from images purely by latent imagination. It learns a latent dynamics world model (RSSM) from past experience, then trains an actor-critic entirely on imagined trajectories in the compact latent space, propagating analytic value gradients back through the learned dynamics via reparameterization. On 20 visual control tasks from the DeepMind Control Suite, Dreamer surpasses prior model-based and model-free agents in data-efficiency, computation time, and final performance, and also transfers to discrete-action Atari and DeepMind Lab tasks."},{"id":"arxiv:1805.09042","kind":"paper","n":156,"label":"Generalisation of structural knowledge in the hippocampal-entorhinal system","authors":"James C.R. Whittington, Timothy H. Muller, Shirley Mark, Caswell Barry, Timothy ","year":2018,"url":"https://arxiv.org/abs/1805.09042","region":"t28","hemi":"llm","lobe":"frontal","color":"#7667b0","indegree":5,"alsoIn":[],"x":0.5613,"y":0.4133,"text":"The paper addresses how structural knowledge can be generalised across environments, taking inspiration from the hippocampal-entorhinal system. It proposes separating representations of world structure (grid cells) from sensory entities, linking them via a conjunctive place-cell code stored in fast Hebbian memory, and trains an unsupervised generative temporal model (VAE with BPTT) to predict sensory observations while an agent walks 2D graph worlds. The network learns brain-like spatial representations (grid, band, border, object-vector, place cells) that emerge without hard-coding, enabling one-shot learning and zero-shot inference of unobserved graph links. Analysis of rat recording data confirms a preserved grid\u2013place cell relationship across environments, supporting non-random place cell remapping."},{"id":"arxiv:2310.01361","kind":"paper","n":157,"label":"GenSim: Generating Robotic Simulation Tasks via Large Language Models","authors":"Wang, Ling, Wang, Yang, Wu, Fan, Wang, Zhang, Liu, Gan","year":2023,"url":"https://arxiv.org/abs/2310.01361","region":"t0","hemi":"wm","lobe":"frontal","color":"#9a4f4f","indegree":0,"alsoIn":[],"x":0.1745,"y":0.6867,"text":"Collecting real-world robot interaction data is expensive, and existing simulation data-generation methods focus on scene-level diversity rather than task-level diversity because authoring novel tasks requires heavy human effort. GenSim uses an LLM's grounding and coding ability to automatically generate simulation environments, task code, and expert demonstrations in two modes: goal-directed (LLM proposes a task curriculum toward a target task) and exploratory (LLM bootstraps from prior tasks to propose novel ones). Using GPT-4, the authors expand the 10-task Ravens/CLIPort benchmark to over 100 tasks, finetune GPT-3.5 and Code-Llama on the generated library, and train language-conditioned multitask policies. Policies pretrained on GPT-4-generated tasks improve task-level generalization (>50% in-domain, ~40% zero-shot in sim) and transfer to unseen long-horizon real-world tasks, outperforming baselines by 25%."},{"id":"arxiv:2401.13178","kind":"paper","n":158,"label":"AgentBoard: An Analytical Evaluation Board of Multi-turn LLM Agents","authors":"Chang Ma et al.","year":2024,"url":"https://arxiv.org/abs/2401.13178","region":"t35","hemi":"llm","lobe":"occipital","color":"#b0679e","indegree":4,"alsoIn":[],"x":0.6251,"y":0.9527,"text":"AgentBoard addresses the problem that existing LLM-agent benchmarks lack task diversity, multi-round interaction, partially-observable environments, and process-level metrics, relying mainly on final success rate that fails to distinguish models scoring near zero. It introduces a benchmark of 9 diverse tasks (1013 environments) across embodied, game, web, and tool scenarios, unified under a multi-round reflex agent, plus a fine-grained progress rate metric (via manually annotated subgoals or matching scores) and an open-source analytical evaluation toolkit with WandB visualization. Evaluating 17+ proprietary and open-weight LLMs, GPT-4 leads with 70.0% avg progress rate / 47.9% success rate, and analyses reveal that progress rate discriminates models that success rate cannot, and that agentic ability depends on grounding, world modeling, self-reflection, and code training. The progress rate correlates with human judgment at Pearson >0.95 on all tasks."},{"id":"arxiv:2112.07103","kind":"paper","n":159,"label":"Hierarchical Stochastic Scheduling of Multi-Community Integrated Energy Systems in Uncertain Environments via Stackelberg Game","authors":"","year":null,"url":"https://arxiv.org/abs/2112.07103","region":"t4","hemi":"wm","lobe":"frontal","color":"#9a7c4f","indegree":6,"alsoIn":[],"x":0.3092,"y":0.1424,"text":"This paper addresses energy management and pricing in multi-community integrated energy systems (MCIESs) with renewable uncertainty. It uses WGAN-GP to generate wind/photovoltaic scenarios, Kmeans++ for scenario reduction, and a Stackelberg game where the MCIES operator (leader) sets energy prices to maximize profit while building users (followers) adjust consumption to minimize costs. Results on a real North China MCIES show that integrated demand response and multi-energy interaction achieve a win-win: operator net profit of \u00a511,981 with reduced user costs compared to baselines."},{"id":"arxiv:2406.00877","kind":"paper","n":160,"label":"Evidence of Learned Look-Ahead in a Chess-Playing Neural Network","authors":"Jenner et al. (Erik Jenner, Shreyas Kapur, Vasil Georgiev, Cameron Allen, Scott ","year":2024,"url":"https://arxiv.org/abs/2406.00877","region":"t20","hemi":"llm","lobe":"frontal","color":"#67b0b0","indegree":2,"alsoIn":[],"x":0.651,"y":0.7613,"text":"The paper investigates whether neural networks learn algorithms like look-ahead/search 'in the wild' rather than relying purely on heuristics, using the policy network of Leela Chess Zero (a transformer that treats each of the 64 board squares as a token). Through mechanistic interpretability on a filtered dataset of 22.5k hard Lichess tactics puzzles, the authors present three converging lines of evidence: activation patching shows the target square of the optimal 3rd move (two turns ahead) is unusually causally important; specific attention heads move information 'backward in time' (L12H12) and 'forward in time' (piece-movement heads); and a simple bilinear probe predicts the 3rd move target square with 92% accuracy. The findings constitute an existence proof that networks can internally represent and causally use future optimal moves, i.e., learned look-ahead."},{"id":"nature.com:e56ca063b728","kind":"paper","n":161,"label":"Mastering the Game of Go with Deep Neural Networks and Tree Search","authors":"David Silver, Aja Huang, Christopher J. Maddison, Arthur Guez, Laurent Sifre, Ge","year":2016,"url":"https://www.nature.com/articles/nature16961","region":"t22","hemi":"llm","lobe":"temporal","color":"#4f849a","indegree":2,"alsoIn":[],"x":0.6286,"y":0.7685,"text":"The paper tackles computer Go, long considered AI's hardest classic board game due to its enormous search space (~250^150) and the difficulty of evaluating positions. It introduces AlphaGo, which combines deep convolutional 'policy networks' (to select moves) and 'value networks' (to evaluate positions) \u2014 trained by supervised learning from human expert games then refined by reinforcement learning through self-play \u2014 with a new Monte Carlo tree search (MCTS) algorithm that fuses these networks with rollouts. AlphaGo achieved a 99.8% win rate against other Go programs and defeated the human European champion Fan Hui 5 games to 0, the first time a program beat a professional human on the full 19x19 board with no handicap."},{"id":"arxiv:2310.08803","kind":"paper","n":162,"label":"Advancing Perception in Artificial Intelligence through Principles of Cognitive Science","authors":"","year":null,"url":"https://arxiv.org/abs/2310.08803","region":"t5","hemi":"llm","lobe":"parietal","color":"#b09e67","indegree":4,"alsoIn":[],"x":0.6719,"y":0.5439,"text":"This review paper maps the cognitive-science of human perception onto current AI techniques, breaking perception into five processes (sensory stimulation, modularity/multisensory integration, bottom-up processing, top-down processing, and interpretation) and, for each, surveying theories from neuroscience, psychology, and linguistics alongside their AI counterparts (Table I). For every cognitive theory it draws a parallel to an existing AI method\u2014e.g. retinotopicity/cortical magnification to fovea-inspired CNN sampling, the mental lexicon to retrieval-augmented/knowledge-base language models, predictive coding to backpropagation and spiking networks, and dual-process theory to heuristic-plus-systematic reasoning. The authors then catalog concrete research gaps where AI has no or only crude analogues (temporal-coincidence dynamic grouping, object permanence, the phonological buffer, dynamic tokenization vocabularies, progressive modularization). They conclude with a five-element schema for general intelligence (functional/cognitive knowledge, generalization, embodied cognition, interpretability, social intelligence) and argue against narrowly task-specific designs."},{"id":"arxiv:2505.02074","kind":"paper","n":163,"label":"Learning Local Causal World Models with State Space Models and Attention","authors":"","year":null,"url":"https://arxiv.org/abs/2505.02074","region":"t6","hemi":"wm","lobe":"temporal","color":"#b0a967","indegree":5,"alsoIn":[],"x":0.1815,"y":0.359,"text":"The paper tackles whether State Space Model (SSM) architectures can learn causal world models, not just predict dynamics, since causal representations are argued to be necessary for robust agents. It proposes S2-SSM (Sparse Slot State Space Model), which augments a SlotSSM/Mamba object-centric video world model with attention-weight-derived adjacency matrices and a path-counting sparsity regularization (adapted from the SPARTAN Transformer) to discover a local causal graph over object/environment slots. On the synthetic Interventional Pong dataset (32\u00d732 videos, 11 environments), S2-SSM matches or beats an equivalent sparse Transformer (S2-TE) on both next-frame reconstruction (MSE) and causal-graph correctness (SHD). An ablation shows that removing the sparsity regularization barely affects reconstruction but destroys the causal graph, which collapses to connecting all object pairs."},{"id":"arxiv:2401.07851","kind":"paper","n":164,"label":"Unlocking Efficiency in Large Language Model Inference: A Comprehensive Survey of Speculative Decoding","authors":"Heming Xia, Zhe Yang, Qingxiu Dong, et al.","year":2024,"url":"https://arxiv.org/abs/2401.07851","region":"t9","hemi":"llm","lobe":"parietal","color":"#97b067","indegree":7,"alsoIn":[],"x":0.875,"y":0.7691,"text":"This paper is the first comprehensive survey of Speculative Decoding, a Draft-then-Verify paradigm that accelerates autoregressive LLM inference by efficiently drafting multiple future tokens and verifying them in parallel with the target LLM. It provides a formal definition and formulation, a taxonomy organizing methods by drafting strategy (independent vs. self-drafting) and verification criterion (greedy, speculative sampling, token tree verification), and a review of alignment/knowledge-distillation techniques. The authors introduce Spec-Bench, a benchmark spanning six tasks, and run a third-party comparative evaluation showing EAGLE achieves the highest overall speedup (~2.0\u20132.5\u00d7) under unified conditions. The survey concludes with challenges around batched inference, drafter efficiency trade-offs, and integration with other efficiency techniques."},{"id":"arxiv:2309.17382","kind":"paper","n":165,"label":"Reason for Future, Act for Now: A Principled Framework for Autonomous LLM Agents with Provable Sample Efficiency","authors":"Liu et al.","year":2023,"url":"https://arxiv.org/abs/2309.17382","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":1,"alsoIn":[],"x":0.7273,"y":0.8289,"text":"The paper tackles the lack of a rigorous mapping between reinforcement learning and LLM reasoning that would let autonomous LLM agents complete tasks within a provably minimal number of environment interactions. It proposes RAFA ('reason for future, act for now'), which casts LLM reasoning as learning and planning in a Bayesian adaptive MDP: a learning subroutine prompts LLMs (Model and Critic) to estimate the environment from a memory buffer, a planning subroutine generates a long-horizon optimal trajectory (tree-search/MCTS/value iteration), the agent executes only the first action, then replans from the new state. Theoretically it proves a \u221aT Bayesian regret bound (e.g. \u00d5((1-\u03b3)\u207b\u00b9(\u03ba+1)\u221a(d\u00b3T)) for Bayesian linear kernel MDPs), with \u03ba-free variants via optimistic bonus and posterior sampling. Empirically RAFA achieves SOTA on Game of 24, ALFWorld, BlocksWorld, and a new Tic-Tac-Toe benchmark."},{"id":"arxiv:2010.11929","kind":"paper","n":166,"label":"An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale","authors":"Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua","year":2020,"url":"https://arxiv.org/abs/2010.11929","region":"t23","hemi":"llm","lobe":"occipital","color":"#4f799a","indegree":0,"alsoIn":[],"x":0.8051,"y":0.2259,"text":"The paper challenges the assumption that convolutional networks are necessary for image recognition by applying a near-standard Transformer encoder directly to sequences of fixed-size image patches (16x16) treated as tokens. When pre-trained on large datasets (ImageNet-21k with 14M images or the in-house JFT-300M with 303M images) and transferred to downstream benchmarks, the Vision Transformer (ViT) matches or exceeds state-of-the-art CNNs while requiring substantially less pre-training compute. The key finding is that large-scale pre-training compensates for ViT's lack of image-specific inductive biases (locality, translation equivariance): on small datasets ViT underperforms ResNets, but it overtakes them as data scales."},{"id":"arxiv:2506.23068","kind":"paper","n":167,"label":"Curious Causality-Seeking Agents Learn Meta Causal World","authors":"et al.","year":2025,"url":"https://arxiv.org/abs/2506.23068","region":"t6","hemi":"wm","lobe":"temporal","color":"#b0a967","indegree":3,"alsoIn":[],"x":0.19,"y":0.2926,"text":"The paper argues that in open-ended environments the apparent causal mechanism drifts because agents observe only local, context-limited windows, so a single fixed causal graph misleads world models. It introduces the Meta-Causal Graph (MCG), a unified representation composed of multiple causal subgraphs each triggered by a latent 'meta state', and a Curious Causality-Seeking Agent that uses curiosity-driven interventions to identify meta states, orient edges, and iteratively refine the graph. Theoretically it proves identifiability of meta states (up to label-swap / observational equivalence) and of causal subgraphs under suitable interventions. Empirically MCG beats observational and prior causal baselines on the Chemical benchmark and a Robosuite Magnetic robot-arm task, e.g. 63.18% OOD prediction accuracy on Fork n=2 vs 57.82% for the best baseline FCDL."},{"id":"arxiv:2508.13073","kind":"paper","n":168,"label":"Large VLM-based Vision-Language-Action Models for Robotic Manipulation: A Survey","authors":"Multiple","year":2025,"url":"https://arxiv.org/abs/2508.13073","region":"t0","hemi":"llm","lobe":"frontal","color":"#9a4f4f","indegree":13,"alsoIn":[],"x":0.5717,"y":0.2084,"text":"This survey provides the first systematic, taxonomy-oriented review of large VLM-based Vision-Language-Action (VLA) models for robotic manipulation. It proposes a two-part taxonomy\u2014Monolithic (single-system and dual-system) and Hierarchical (planner-only and planner+policy)\u2014and synthesizes advances across architecture design, RL optimization, training-free methods, learning from human videos, and world model integration. The paper also catalogs datasets, benchmarks, and defining characteristics such as multimodal fusion, instruction following, and multi-dimensional generalization, while identifying future directions including memory mechanisms, 4D perception, and multi-agent cooperation."},{"id":"arxiv:2406.16062","kind":"paper","n":169,"label":"Towards Biologically Plausible Computing: A Comprehensive Comparison","authors":"Changze Lv, Yufei Gu, Zhengkang Guo, et al.","year":2024,"url":"https://arxiv.org/abs/2406.16062","region":"t20","hemi":"llm","lobe":"frontal","color":"#67b0b0","indegree":4,"alsoIn":[],"x":0.6046,"y":0.6228,"text":"The paper addresses the long-standing question of whether backpropagation is biologically plausible by first establishing five criteria for biological plausibility (weight asymmetry, local error, non-parallel training, spiking neuron models, unsigned errors), then reviewing and empirically comparing nine brain-inspired learning algorithms (Hebbian learning, STDP, feedback alignment, target propagation, predictive coding, forward-forward, perturbation learning, local losses, energy-based/equilibrium propagation). The algorithms are evaluated on image classification (MNIST, CIFAR-10, CIFAR-100) with MLP and CNN architectures, and their learned representations are compared to human fMRI brain activity (Haxby dataset) via representational similarity analysis. Results show all bio-plausible methods trail backpropagation on accuracy, with local losses closest, while predictive coding and forward-forward achieve the highest representational similarity to brain activity, revealing an accuracy-vs-plausibility trade-off."},{"id":"arxiv:2212.09748","kind":"paper","n":170,"label":"Scalable Diffusion Models with Transformers","authors":"William Peebles, Saining Xie","year":2023,"url":"https://arxiv.org/abs/2212.09748","region":"t32","hemi":"wm","lobe":"frontal","color":"#8b4f9a","indegree":2,"alsoIn":[],"x":0.387,"y":0.2013,"text":"The paper investigates whether the convolutional U-Net backbone, the de-facto standard for diffusion models, is essential, and replaces it with a pure transformer operating on latent image patches (Diffusion Transformers, DiTs) within the Latent Diffusion Model framework. It analyzes scalability through forward-pass complexity (Gflops) and finds that increasing transformer depth/width or decreasing patch size (more tokens) consistently lowers FID. The largest model, DiT-XL/2, outperforms all prior diffusion models, achieving a state-of-the-art FID of 2.27 on class-conditional ImageNet 256\u00d7256 and 3.04 on 512\u00d7512. The work shows the U-Net inductive bias is not crucial and that diffusion models inherit transformer scaling properties."},{"id":"arxiv:2401.02954","kind":"paper","n":171,"label":"DeepSeek LLM: Scaling Open-Source Language Models with Longtermism","authors":"Xiao Bi et al. (DeepSeek)","year":2024,"url":"https://arxiv.org/abs/2401.02954","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":9,"alsoIn":[],"x":0.8021,"y":0.5909,"text":"DeepSeek LLM studies scaling laws to guide training of open-source 7B and 67B bilingual (Chinese/English) language models on a 2-trillion-token corpus, then aligns them via SFT and DPO. The authors derive power-law scaling formulae for batch size and learning rate vs compute budget, introduce non-embedding FLOPs/token (M) as a more accurate model-scale representation, and show that data quality shifts the optimal model/data allocation. DeepSeek LLM 67B Base outperforms LLaMA-2 70B across code, math, and reasoning benchmarks, and DeepSeek 67B Chat surpasses GPT-3.5 in open-ended Chinese (AlignBench) and English (MT-Bench) evaluations."},{"id":"arxiv:2310.06114","kind":"paper","n":172,"label":"Learning Interactive Real-World Simulators (UniSim)","authors":"Yang, Du, Ghasemipour, Tompson, Schuurmans, Abbeel et al.","year":2023,"url":"https://arxiv.org/abs/2310.06114","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":4,"alsoIn":[],"x":0.2484,"y":0.6952,"text":"The paper introduces UniSim, a universal action-conditioned video generation model that learns an interactive simulator of the real world by orchestrating diverse datasets (internet text-image, human activity, robot, navigation, and simulated data) under a unified action-in-video-out interface. It is formulated as an observation-prediction model parametrized by a 5.6B-parameter video diffusion (3D U-Net) that can be rolled out autoregressively for long-horizon, consistent simulation from high-level language instructions or low-level motor controls. The authors show UniSim can train high-level vision-language policies (via hindsight relabeling), low-level RL control policies (via model-based REINFORCE), and video captioning VLMs purely in simulation, with policies transferring zero-shot to real robots. Results demonstrate large gains in long-horizon task completion, RL success rates, and captioning transfer over baselines trained on real/original data."},{"id":"arxiv:2206.10498","kind":"paper","n":173,"label":"PlanBench: An Extensible Benchmark for Evaluating Large Language Models on Planning and Reasoning about Change","authors":"Valmeekam, Marquez, Olmo, Sreedharan & Kambhampati","year":2023,"url":"https://arxiv.org/abs/2206.10498","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":0,"alsoIn":[],"x":0.7318,"y":0.8424,"text":"The paper introduces PlanBench, an extensible benchmark for systematically evaluating whether LLMs can plan and reason about actions and change, grounded in IPC-style domains (Blocksworld, Logistics) rather than commonsense tasks where retrieval is hard to distinguish from planning. It defines eight test cases (plan generation, cost-optimal planning, plan verification, reasoning about plan execution, robustness to goal reformulation, plan reuse, replanning, plan generalization) auto-generated via PDDL and validated mechanistically with a planner (Fast-Downward) and validator (VAL). Evaluating GPT-4 and InstructGPT-3 (text-davinci-002), the authors find LLMs perform poorly on core planning tasks \u2014 e.g., GPT-4 generates valid plans only 34.3% of the time in Blocksworld. Performance collapses further under domain obfuscation, suggesting reliance on pattern matching from background knowledge rather than genuine reasoning."},{"id":"arxiv:2305.17926","kind":"paper","n":174,"label":"Large Language Models are not Fair Evaluators","authors":"Peiyi Wang et al.","year":2023,"url":"https://arxiv.org/abs/2305.17926","region":"t33","hemi":"llm","lobe":"parietal","color":"#ad67b0","indegree":4,"alsoIn":[],"x":0.8948,"y":0.6994,"text":"The paper shows that using LLMs (GPT-4, ChatGPT) as referees to score and compare candidate model responses suffers from a systematic positional bias: simply swapping the order of two responses in the prompt can reverse the evaluation outcome, so much so that Vicuna-13B could be made to beat ChatGPT on 66/80 queries with ChatGPT as judge. To fix this, the authors propose a calibration framework of three strategies \u2014 Multiple Evidence Calibration (generate reasoning before scoring, sample k times), Balanced Position Calibration (average scores across both response orders), and Human-in-the-Loop Calibration (a Balanced Position Diversity Entropy to flag hard cases for human review). On 80 manually annotated Vicuna-benchmark win/tie/lose examples, MEC+BPC raises alignment accuracy and kappa with humans, and HITLC with 20% human effort reaches near-human accuracy while cutting annotation cost ~39%."},{"id":"arxiv:2606.20545","kind":"paper","n":175,"label":"Current World Models Lack a Persistent State Core","authors":"Jinpeng Lu, Dexu Zhu, Haoyuan Shi, Linghan Cai, Guo Tang, Yinda Chen, Jie Cao, D","year":2026,"url":"https://arxiv.org/abs/2606.20545","region":"t17","hemi":"llm","lobe":"parietal","color":"#4f9a79","indegree":0,"alsoIn":[],"x":0.648,"y":0.289,"text":"The paper argues that current video generators marketed as world models lack a persistent state core: they fail to keep the world evolving while it is unobserved, instead resuming a returning target in the state it was abandoned. It introduces WRBench, a diagnostic benchmark that treats camera motion as an intervention on observability and scores six hierarchical dimensions (from requested-camera precision through re-observed spatial/state consistency), grounded in the Natural-25 prompt suite and the WRBenchLib provenance toolkit. Evaluating 23 models on 9,600 videos (calibrated by 2,547 human verdicts), it finds a preservation\u2013access\u2013re-observed-consistency gap where visible fidelity, camera control, geometric priors, and parameter scale all fail to bind the hidden event endpoint. It concludes world models need a 'what-memory' that records hidden change plus an endpoint-persistence training objective."},{"id":"arxiv:2405.10480","kind":"paper","n":176,"label":"Lean Attention: Hardware-Aware Scalable Attention Mechanism for the Decode-Phase of Transformers","authors":"","year":null,"url":"https://arxiv.org/abs/2405.10480","region":"t16","hemi":"llm","lobe":"frontal","color":"#67b084","indegree":3,"alsoIn":[],"x":0.8185,"y":0.2903,"text":"Transformer decode-phase inference is memory-bandwidth bound and existing attention kernels (FlashAttention-2, FlashDecoding) underutilize GPU compute at long context lengths because they cannot balance work across streaming multiprocessors. LeanAttention re-frames the online-softmax re-scaling as an associative reduction operation, letting attention be partitioned along the context-length dimension using a stream-K-style equalized decomposition into 'LeanTiles' executed in a single fused kernel. This yields near-100% SM occupancy across problem sizes and tensor-parallel multi-GPU scaling. It achieves an average 2.6x attention speedup over FlashAttention-2 and up to 8.33x at 512k context length."},{"id":"arxiv:2203.03417","kind":"paper","n":177,"label":"Scalable multi-agent reinforcement learning for distributed control of residential energy flexibility","authors":"","year":null,"url":"https://arxiv.org/abs/2203.03417","region":"t4","hemi":"wm","lobe":"frontal","color":"#9a7c4f","indegree":2,"alsoIn":[],"x":0.1548,"y":0.2744,"text":"The paper addresses the scalability problem of coordinating residential energy flexibility (EVs, space heating, flexible loads) using multi-agent reinforcement learning (MARL) in partially observable stochastic environments. Standard independent Q-learning loses coordination performance at scale, so the authors propose a novel combination of learning from off-line convex optimizations on historical data and using marginal reward signals that isolate each agent's contribution to global rewards, all with fixed-size Q-tables for privacy-preserving decentralized control. The best-performing strategy (MO) maintains stable performance up to 30 agents, achieving 33.7% cost reductions per agent without sharing personal data."},{"id":"arxiv:2205.06175","kind":"paper","n":178,"label":"A Generalist Agent","authors":"Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander ","year":2022,"url":"https://arxiv.org/abs/2205.06175","region":"t0","hemi":"llm","lobe":"frontal","color":"#9a4f4f","indegree":0,"alsoIn":[],"x":0.6819,"y":0.6659,"text":"Gato is a single multi-modal, multi-task, multi-embodiment generalist agent instantiated as a 1.2B-parameter decoder-only transformer that serializes images, text, proprioception, and continuous/discrete actions into a flat token sequence and trains autoregressively (supervised behavior cloning) on 604 tasks across 596 control datasets plus vision-language data. With a single set of weights it can play Atari, caption images, chat, navigate 3D environments, and stack blocks with a real robot arm, choosing its output modality from context. It performs above 50% of expert score on over 450 of 604 tasks, reaches average-human-or-better Atari scores on 23 games, and stacks blocks on a real robot competitively with a published BC baseline. Scaling analysis across 79M/364M/1.18B models shows consistent in-distribution gains with model size, and fine-tuning on as few as 10 demonstrations recovers expert robot performance."},{"id":"arxiv:2411.06559","kind":"paper","n":179,"label":"Is Your LLM Secretly a World Model of the Internet? Model-Based Planning for Web Agents","authors":"Gu et al.","year":2024,"url":"https://arxiv.org/abs/2411.06559","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":3,"alsoIn":[],"x":0.6241,"y":0.9184,"text":"Web agents face irreversible actions on real websites, making tree-search backtracking infeasible and slow. The paper proposes WEBDREAMER, a model-based planning framework that uses an LLM as both a world model (simulating natural-language state changes of candidate actions) and a value/scoring function, executing the highest-scoring action via Model Predictive Control. It also trains a specialized 7B world model (Dreamer-7B) on 3.1M synthesized web-interaction instances. WEBDREAMER beats reactive baselines on three benchmarks, is competitive with tree search while being 4\u20135\u00d7 more efficient, and Dreamer-7B performs comparably to GPT-4o."},{"id":"arxiv:2508.06297","kind":"paper","n":180,"label":"KV Cache Compression for Inference Efficiency in LLMs: A Review","authors":"","year":null,"url":"https://arxiv.org/abs/2508.06297","region":"t18","hemi":"llm","lobe":"temporal","color":"#67b09a","indegree":7,"alsoIn":[],"x":0.8993,"y":0.4285,"text":"This review surveys KV cache compression techniques for improving LLM inference efficiency, motivated by the linear growth of KV cache memory with sequence length and batch size that creates a serving bottleneck as contexts reach 128K-to-million tokens. It taxonomizes ~25 published methods into four families\u2014selective token compression, quantization, attention compression, and hybrid approaches\u2014and tabulates each method's reported throughput, inference-efficiency, compression-ratio, and perplexity figures, mostly on LLaMa-family models. The authors then compare methods head-to-head on LLaMa (e.g., NACL gives the highest inference-rate gain at 78%, CacheBlend/DistAttention give the largest throughput gains at 3.9\u00d7 and 3.61\u00d7) and identify gaps in cross-stage collaboration, dynamic adaptation, and hardware-software co-design. They close with three future directions: hybrid optimization/integration, adaptive dynamic optimization, and hardware-software co-optimization."},{"id":"arxiv:2310.08864","kind":"paper","n":181,"label":"Open X-Embodiment: Robotic Learning Datasets and RT-X Models","authors":"Open X-Embodiment Collaboration (Padalkar, Pooley, Jain et al.)","year":2023,"url":"https://arxiv.org/abs/2310.08864","region":"t0","hemi":"wm","lobe":"frontal","color":"#9a4f4f","indegree":1,"alsoIn":[],"x":0.1635,"y":0.7048,"text":"The paper addresses whether a single 'generalist' robot policy can be trained across many different robots and benefit from positive cross-embodiment transfer, rather than training separate models per robot/task/environment. The authors assemble the Open X-Embodiment (OXE) Dataset by pooling 60 existing datasets (34 labs, 21 institutions) into a standardized RLDS format, covering 22 robot embodiments, 1M+ real trajectories, and 527 skills (160,266 tasks). They train two existing architectures\u2014RT-1 (35M params) and RT-2 (a 55B VLM-based VLA)\u2014on this data with minimal modification, producing RT-1-X and RT-2-X. Results show RT-1-X improves small-data domains by ~50% on average over original methods, while the high-capacity RT-2-X shows ~3\u00d7 emergent-skill transfer and improved generalization, demonstrating that co-training on multi-robot data yields positive transfer when model capacity is sufficient."},{"id":"arxiv:2405.12399","kind":"paper","n":182,"label":"Diffusion for World Modeling: Visual Details Matter in Atari (DIAMOND)","authors":"Alonso, Jelley, Micheli, Kanervisto, Storkey, Pearce & Fleuret","year":2024,"url":"https://arxiv.org/abs/2405.12399","region":"t32","hemi":"wm","lobe":"frontal","color":"#8b4f9a","indegree":4,"alsoIn":[],"x":0.1771,"y":0.3994,"text":"DIAMOND (DIffusion As a Model Of eNvironment Dreams) is a reinforcement learning agent trained entirely inside a diffusion-based world model, departing from the dominant discrete-latent paradigm in order to preserve visual details that matter for control. The authors adapt the EDM diffusion formulation (rather than DDPM) to condition next-frame generation on past frames and actions via a U-Net 2D with frame-stacking, achieving stable autoregressive rollouts with as few as 1-3 denoising steps. DIAMOND reaches a mean human-normalized score of 1.46 on Atari 100k, a new best among in-world-model agents, with improvements concentrated in games where small visual details are critical. They further scale the world model to a real-time interactive neural game engine for CS:GO Dust II trained on 87 hours of static gameplay."},{"id":"arxiv:2505.01712","kind":"paper","n":183,"label":"World Model-Based Learning for Long-Term Age of Information Minimization in Vehicular Networks","authors":"","year":null,"url":"https://arxiv.org/abs/2505.01712","region":"t39","hemi":"wm","lobe":"occipital","color":"#9a4f5b","indegree":3,"alsoIn":[],"x":0.1995,"y":0.5296,"text":"Traditional RL approaches (model-free and model-based) for wireless networks suffer from low data efficiency, short-sighted policies, and inability to handle long-term credit assignment in highly dynamic mmWave V2X networks. This paper proposes a world model-based learning framework that combines a recurrent state-space model (RSSM) with an actor-critic policy to learn link scheduling in differentiable imagined latent trajectories, minimizing packet-completeness-aware age of information (CAoI). The world model's imagination ability also lets it jointly predict time-varying wireless data and schedule links during intervals without real observations. Evaluated on a Sionna-based physics-accurate mmWave V2X simulator, it achieves large data-efficiency gains and improves CAoI by 26% over MBRL and 16% over MFRL baselines."},{"id":"arxiv:2206.14176","kind":"paper","n":184,"label":"DayDreamer: World Models for Physical Robot Learning","authors":"Wu, Escontrela, Hafner, Goldberg, Abbeel","year":2022,"url":"https://arxiv.org/abs/2206.14176","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":2,"alsoIn":[],"x":0.1797,"y":0.4841,"text":"DayDreamer applies the Dreamer world-model algorithm (DreamerV2) to physical robot learning, training 4 real robots online from scratch without simulators or demonstrations. A world model learns latent dynamics from a replay buffer, and an actor-critic learns behaviors by imagined rollouts in latent space, with decoupled async data collection and learning. Using identical hyperparameters across all robots, Dreamer teaches a quadruped to roll over, stand, and walk in 1 hour, learns visual pick-and-place on two arms in 8-10 hours approaching human performance, and learns visual navigation on a wheeled robot in under 2 hours."},{"id":"arxiv:2603.20397","kind":"paper","n":185,"label":"KV Cache Optimization Strategies for Scalable and Efficient LLM Inference","authors":"","year":null,"url":"https://arxiv.org/abs/2603.20397","region":"t18","hemi":"llm","lobe":"temporal","color":"#67b09a","indegree":5,"alsoIn":[],"x":0.9292,"y":0.4798,"text":"This survey systematically reviews recent KV cache optimization techniques for Transformer-based LLM inference, where the cache's memory footprint grows linearly with context length and bottlenecks GPU memory, bandwidth, and throughput as context windows scale to millions of tokens. The authors organize methods into five categories\u2014cache eviction, cache compression, hybrid memory solutions, novel attention mechanisms, and combination strategies\u2014analyzing mechanisms, trade-offs, and reported empirical performance (memory reduction, throughput, accuracy) for each. They further map techniques onto seven practical deployment scenarios (long-context single requests, high-throughput serving, edge devices, multi-turn conversations, prefill-heavy, accuracy-critical reasoning, hardware-specific). The core conclusion is that no single technique dominates across all settings; optimal choice depends on context length, hardware, and workload, motivating adaptive multi-stage optimization pipelines."},{"id":"arxiv:2408.14837","kind":"paper","n":186,"label":"Diffusion Models Are Real-Time Game Engines (GameNGen)","authors":"Valevski et al.","year":2024,"url":"https://arxiv.org/abs/2408.14837","region":"t32","hemi":"wm","lobe":"frontal","color":"#8b4f9a","indegree":3,"alsoIn":[],"x":0.2781,"y":0.4272,"text":"GameNGen is the first game engine powered entirely by a neural model that enables real-time interactive simulation of a complex game (DOOM) over long trajectories. It works in two phases: an RL (PPO) agent plays the game to generate recorded training trajectories, then an augmented Stable Diffusion v1.4 diffusion model is trained to predict the next frame conditioned on past frames and actions. The system runs at 20 FPS on a single TPU, achieves next-frame PSNR of 29.4 (comparable to lossy JPEG), and human raters distinguish real game clips from simulation only slightly above chance even after 5 minutes of autoregressive play. Key technical insights are noise augmentation of context frames to prevent autoregressive drift and latent-decoder fine-tuning to improve detail/HUD fidelity."},{"id":"arxiv:2401.09985","kind":"paper","n":187,"label":"WorldDreamer: Towards General World Models for Video Generation via Predicting Masked Tokens","authors":"Wang, Zhu, Huang et al.","year":2024,"url":"https://arxiv.org/abs/2401.09985","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":5,"alsoIn":[],"x":0.387,"y":0.6757,"text":"World models have been confined to narrow domains (gaming, driving, robotics), limiting their ability to capture general-world dynamics for video generation. WorldDreamer reframes general world modeling as unsupervised visual sequence modeling: it tokenizes images/videos with VQGAN, randomly masks tokens, and trains a Transformer to predict the masked tokens conditioned on unmasked tokens plus multimodal (text + action) prompts. A proposed Spatial Temporal Patchwise Transformer (STPT) restricts attention to local spatial-temporal patches to speed convergence, and parallel masked-token decoding produces a 24-frame 192\u00d7320 video in ~3s on a single A800. The model handles text-to-video, image-to-video, video inpainting, stylization, and action-to-video across natural and driving scenes, ~3\u00d7\u201320\u00d7 faster than diffusion/autoregressive baselines."},{"id":"arxiv:2503.14492","kind":"paper","n":188,"label":"Cosmos-Transfer1: Conditional World Generation with Adaptive Multimodal Control","authors":"NVIDIA (Hassan Abu Alhaija et al.)","year":2025,"url":"https://arxiv.org/abs/2503.14492","region":"t15","hemi":"wm","lobe":"occipital","color":"#4f9a62","indegree":0,"alsoIn":[],"x":0.3986,"y":0.845,"text":"Cosmos-Transfer1 is a diffusion-based conditional world generation model that synthesizes simulation videos from multiple spatial control modalities (segmentation, depth, Canny edge, blur visual, and\u2014for driving\u2014HDMap and LiDAR). It extends the Cosmos-Predict1-7B DiT world model with one ControlNet branch per modality, trained separately and fused at inference via an adaptive spatiotemporal control map that weights each modality differently per location and time. Evaluations on the curated 600-example TransferBench show the multimodal model achieves the best overall quality while preserving scene structure, and case studies demonstrate robotics Sim2Real and autonomous-vehicle data enrichment. An inference-scaling strategy on an NVIDIA GB200 NVL72 rack reaches real-time generation (5-second video in 4.2 s on 64 GPUs)."},{"id":"arxiv:2007.08794","kind":"paper","n":189,"label":"Discovering Reinforcement Learning Algorithms","authors":"Oh et al.","year":2020,"url":"https://arxiv.org/abs/2007.08794","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":7,"alsoIn":[],"x":0.1683,"y":0.309,"text":"Designing RL update rules has historically required years of manual research; this paper asks whether an entire update rule \u2014 including 'what to predict' (e.g. value functions) and 'how to learn from it' (e.g. bootstrapping) \u2014 can be discovered automatically from data. The authors introduce Learned Policy Gradient (LPG), a backward-LSTM meta-learner that, by interacting with a distribution of toy environments, outputs targets for an agent's policy and a 30-dimensional prediction vector whose semantics are not enforced but discovered. Empirically LPG discovers its own alternative to value functions plus a bootstrapping mechanism, and analysis shows its predictions can recover true values at multiple discount factors better than a scalar TD baseline. Most strikingly, an LPG meta-trained only on tabular/toy domains generalises to complex Atari games, achieving super-human performance on 14 of the 57 games."},{"id":"arxiv:2210.04133","kind":"paper","n":190,"label":"Adapting Pretrained Vision-Language Foundational Models to Medical Imaging Domains","authors":"","year":null,"url":"https://arxiv.org/abs/2210.04133","region":"t13","hemi":"llm","lobe":"parietal","color":"#6bb067","indegree":6,"alsoIn":[],"x":0.6625,"y":0.1031,"text":"The paper investigates whether the pretrained Stable Diffusion vision-language foundation model can be adapted to generate medical chest X-ray (CXR) images despite never being trained on medical data. The authors probe each sub-component (VAE, CLIP text encoder, U-Net) and compare fine-tuning strategies: textual projection (swapping CLIP for an in-domain text encoder with a trained projection), textual inversion, and U-Net fine-tuning. They find the frozen VAE reconstructs CXRs out-of-the-box and the frozen CLIP encoder represents radiology prompts surprisingly well, while U-Net fine-tuning with a prior is the only method that produces high-fidelity CXRs able to insert a realistic pleural effusion. Their best model maintains 95% accuracy on a DenseNet-121 classifier trained to detect the abnormality."},{"id":"arxiv:2506.21876","kind":"paper","n":191,"label":"Do Vision-Language Models Have Internal World Models? Towards an Atomic Evaluation","authors":"Qiyue Gao, Xinyu Pi, Kevin Liu, Junrong Chen, Ruolan Yang, Xinqi Huang, Xinyu Fa","year":2025,"url":"https://arxiv.org/abs/2506.21876","region":"t17","hemi":"llm","lobe":"parietal","color":"#4f9a79","indegree":1,"alsoIn":[],"x":0.639,"y":0.2769,"text":"The paper investigates whether large Vision-Language Models (VLMs) function as internal world models by proposing a cognitively-inspired two-stage framework that separates perception (visual, spatial, temporal, quantitative, motion) from prediction (mechanistic simulation, transitive inference, compositional inference). The authors build WM-ABench, a benchmark of 23 fine-grained dimensions across 6 simulators (ThreeDWorld, ManiSkill, Habitat 2.0, Physion, Carla) with over 100,000 instances using controlled counterfactual simulations. Running 660 experiments on 15 commercial and open-source VLMs, they find striking limitations: near-random motion-trajectory discrimination, weak spatial/temporal perception, poor physical causality, and entangled (non-disentangled) representations such as associating color with speed. Even frontier models (o3, Gemini-2.5-Pro, GPT-4.5) reach human parity on static perception but still lag badly on spatial, temporal, and compositional inference."},{"id":"arxiv:2403.08540","kind":"paper","n":192,"label":"Language Models Scale Reliably with Over-Training and on Downstream Tasks","authors":"Samir Yitzhak Gadre et al.","year":2024,"url":"https://arxiv.org/abs/2403.08540","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":5,"alsoIn":[],"x":0.8051,"y":0.6634,"text":"Standard scaling laws study the compute-optimal (Chinchilla) regime and predict next-token loss, but real models are over-trained to cut inference cost and compared on downstream accuracy. The authors build a testbed of 104 models (0.011B\u20136.9B params) trained on C4, RedPajama, and RefinedWeb at token multipliers from 5 to 640, then (i) reparameterize the Chinchilla loss law in terms of compute C and token multiplier M to extrapolate over-trained loss, and (ii) propose an exponential-decay/power law mapping validation loss to average downstream top-1 error. The fitted laws predict the C4 loss of a 1.4B/900B-token (32\u00d7 over-trained) run and a 6.9B/138B compute-optimal run to within 0.7% relative error using 300\u00d7 less compute, and predict average downstream error to within a few percent using 20\u00d7 less compute."},{"id":"arxiv:2112.15242","kind":"paper","n":193,"label":"A free energy principle for generic quantum systems","authors":"","year":null,"url":"https://arxiv.org/abs/2112.15242","region":"t25","hemi":"llm","lobe":"parietal","color":"#6779b0","indegree":1,"alsoIn":[],"x":0.5867,"y":0.8136,"text":"The paper reformulates the variational Free Energy Principle (FEP) \u2014 originally a theory of brain function stating that systems minimize a variational free energy upper bound on surprisal \u2014 within a spacetime-background-free, scale-free quantum information theory. Using quantum reference frames (QRFs) formalized via category-theoretic Channel Theory, holographic screens as Markov blankets, and thermodynamic symmetry breaking, the authors show that generic quantum systems can be treated as observers/agents that minimize Bayesian prediction error. The central result is that, taken to its asymptotic limit, the quantum FEP drives systems from separability toward entanglement and is asymptotically equivalent to the Principle of Unitarity (conservation of information), the most fundamental axiom of quantum theory. They argue biological systems likely exploit quantum coherence as a computational and communication resource, and predict detectable Bell/Leggett-Garg inequality violations in macroscopic biological systems."},{"id":"arxiv:2201.06387","kind":"paper","n":194,"label":"The free energy principle made simpler but not too simple","authors":"Karl Friston, Lancelot Da Costa, Noor Sajid, Conor Heins, Kai Ueltzh\u00f6ffer, Grigo","year":2023,"url":"https://arxiv.org/abs/2201.06387","region":"t25","hemi":"llm","lobe":"parietal","color":"#6779b0","indegree":7,"alsoIn":[],"x":0.6002,"y":0.8167,"text":"A pedagogical re-derivation of the free energy principle (FEP) that starts from a Langevin description of random dynamical systems and arrives at a Bayesian mechanics of sentient behaviour using standard statistical-physics results. The argument proceeds in three steps: (i) a NESS density plus sparse (Helmholtz-decomposed) coupling induces conditional independencies that define a 'particular partition' of states into external, sensory, active and internal sets separated by a Markov blanket; (ii) internal states are shown to parameterise a variational density over external states, so autonomous dynamics become gradient flow on variational free energy (an evidence lower bound), i.e. self-evidencing; (iii) over extended paths, and for low-noise 'precise particles', the action of autonomous paths equals an expected free energy that unifies planning-as-inference, optimal control, expected utility and curiosity. The contribution is conceptual unification and clearer exposition rather than new empirical results, illustrated with simulations of handwriting and epistemic visual foraging."},{"id":"arxiv:2309.00941","kind":"paper","n":195,"label":"Emergent Linear Representations in World Models of Self-Supervised Sequence Models","authors":"Nanda, Lee & Wattenberg","year":2023,"url":"https://arxiv.org/abs/2309.00941","region":"t6","hemi":"llm","lobe":"temporal","color":"#b0a967","indegree":7,"alsoIn":[],"x":0.6435,"y":0.7574,"text":"The paper investigates how OthelloGPT, an autoregressive transformer trained only to predict legal Othello moves, represents the board state internally. Contradicting prior work (Li et al., 2023a) that claimed the world model was non-linear, the authors show the board is linearly encoded when probed relative to the current player (MINE/YOURS/EMPTY) rather than absolute colours (BLACK/WHITE/EMPTY). They confirm causality by steering the model's predictions with simple vector addition of probe directions, and further interpret how empty tiles, flipped tiles, and end-game circuits are computed."},{"id":"arxiv:1911.11641","kind":"paper","n":196,"label":"PIQA: Reasoning about Physical Commonsense in Natural Language","authors":"Bisk, Zellers, Le Bras, Gao, Choi","year":2020,"url":"https://arxiv.org/abs/1911.11641","region":"t6","hemi":"llm","lobe":"temporal","color":"#b0a967","indegree":0,"alsoIn":[],"x":0.7692,"y":0.3328,"text":"The paper introduces PIQA (Physical Interaction: Question Answering), a benchmark for physical commonsense reasoning in natural language, where a model must pick the more sensible of two solutions to a stated physical goal. The dataset of ~21K goal-solution pairs (16K+ train, ~2K dev, ~3K test) is sourced from instructables.com prompts and filtered with the AFLite algorithm to remove annotation artifacts. Large pretrained transformers struggle (RoBERTa-Large best at 77.1% test) while humans reach 94.9%, a gap of ~18-20 absolute points. Error analysis shows models fail on relational/versatile concepts like 'before/after', 'top/bottom', and 'water', motivating learning physical knowledge beyond text alone."},{"id":"philpapers.org:7b684275d20a","kind":"paper","n":197,"label":"Conscious artificial intelligence and biological naturalism","authors":"Anil K. Seth","year":2025,"url":"https://philpapers.org/rec/SETCAI-4","region":"t2","hemi":"llm","lobe":"temporal","color":"#b07d67","indegree":7,"alsoIn":[],"x":0.6046,"y":0.7338,"text":"The provided text is not a paper but a Cloudflare security block page preventing access to philpapers.org. No actual paper content, abstract, methods, results, or data are available for indexing. The title suggests a philosophical work on conscious AI and biological naturalism, but no substantive content was retrieved."},{"id":"transformer-circuits.pub:b904370d3db5","kind":"paper","n":198,"label":"On the Biology of a Large Language Model","authors":"Lindsey et al. (Anthropic)","year":2025,"url":"https://transformer-circuits.pub/2025/attribution-graphs/biology.html","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":1,"alsoIn":[],"x":0.7303,"y":0.6293,"text":"The paper reverse-engineers the internal computations of Claude 3.5 Haiku using 'attribution graphs' built on a cross-layer transcoder (CLT) replacement model with 30 million interpretable features, then validates the discovered circuits with feature-inhibition and feature-swapping interventions. Across ten case studies it shows the model performs genuine multi-step reasoning 'in its head' (Dallas\u2192Texas\u2192Austin), plans rhyming words before writing a line of poetry, uses shared language-agnostic circuits across English/French/Chinese, reuses addition 'lookup-table' features in non-arithmetic contexts, and runs default refusal/hallucination circuits gated by 'known-entity' features. It also mechanistically distinguishes faithful from unfaithful (bullshitting / motivated, backward-reasoning) chain-of-thought, dissects a bomb-making jailbreak, and audits a model finetuned with a hidden reward-model-bias goal. The authors stress these are existence-proof case studies: their graphs give satisfying insight on only about a quarter of attempted prompts and capture only a fraction of the model's mechanisms."},{"id":"deepmind.google:39914de6d3c9","kind":"paper","n":199,"label":"Genie 2: A large-scale foundation world model","authors":"Parker-Holder et al. (Google DeepMind)","year":2024,"url":"https://deepmind.google/blog/genie-2-a-large-scale-foundation-world-model/","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":3,"alsoIn":[],"x":0.3378,"y":0.6632,"text":"Genie 2 is a large-scale foundation world model that generates action-controllable, playable 3D environments from a single prompt image (e.g. one produced by Imagen 3), navigable by a human or AI agent via keyboard and mouse. It is an autoregressive latent diffusion model: video frames pass through an autoencoder, and the resulting latent frames feed a large transformer dynamics model trained with a causal mask, with classifier-free guidance used to improve action controllability. Trained on a large video dataset, it exhibits emergent capabilities \u2014 object interactions, physics (water, smoke, gravity, lighting, reflections), character animation, long-horizon memory, and modeling of other agents/NPCs. It produces consistent worlds for up to a minute (most demos 10-20s) and is demonstrated as a generator of unseen evaluation environments for SIMA embodied agents."},{"id":"arxiv:1807.11819","kind":"paper","n":200,"label":"Cognitive Computational Neuroscience","authors":"Nikolaus Kriegeskorte, Pamela K. Douglas","year":2018,"url":"https://arxiv.org/abs/1807.11819","region":"t5","hemi":"llm","lobe":"parietal","color":"#b09e67","indegree":2,"alsoIn":[],"x":0.6166,"y":0.5472,"text":"This review argues that understanding how cognition is implemented in the brain requires building task-performing computational models that are neurobiologically plausible and testing them against brain and behavioral data. It traces the complementary contributions of cognitive science (top-down decomposition of cognition into algorithms), computational neuroscience (bottom-up modeling of how neurons implement component functions), and AI (combining components into intelligent behavior), and proposes their integration as 'cognitive computational neuroscience.' It reviews bottom-up experiment-to-theory methods (connectivity, decoding, and representational models) and top-down theory-to-experiment methods (neural network and cognitive models), noting that deep neural networks trained on object recognition currently provide the best models of primate ventral-stream/IT representations. It calls for a collaborative culture of shareable tasks, data, models, and tests to build and adjudicate among AI-scale brain-computational models."},{"id":"arxiv:2502.14819","kind":"paper","n":201,"label":"Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models (PLDM)","authors":"Vlad Sobal et al. (incl. Yann LeCun)","year":2025,"url":"https://arxiv.org/abs/2502.14819","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":3,"alsoIn":[],"x":0.2069,"y":0.276,"text":"The paper studies how to learn from reward-free offline trajectories of varying quality, systematically comparing model-free RL (goal-conditioned and zero-shot) against model-based planning on navigation tasks. The authors introduce PLDM (Planning with a Latent Dynamics Model), which trains a latent dynamics model using a JEPA architecture with a VICReg-inspired anti-collapse objective and plans at test time with MPPI/MPC. Across 23 datasets in three navigation environments, they find model-free RL needs large high-quality datasets while PLDM is more data-efficient and generalizes best to unseen layouts and to non-goal-reaching tasks. PLDM is the only method to reach competitive performance across all six tested generalization properties, though it trades off ~4x slower inference and weaker trajectory stitching than HILP/GCIQL."},{"id":"nature.com:f005fae15467","kind":"paper","n":202,"label":"Mastering Diverse Control Tasks through World Models (DreamerV3)","authors":"Hafner, Pasukonis, Ba & Lillicrap","year":2025,"url":"https://www.nature.com/articles/s41586-025-08744-2","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":5,"alsoIn":[],"x":0.1583,"y":0.3651,"text":"DreamerV3 (the third generation of Dreamer) is a general model-based reinforcement-learning algorithm that learns a world model and improves behaviour by imagining future trajectories, using a single fixed hyperparameter configuration across all domains. A set of robustness techniques based on normalization, balancing, and signal transformations (symlog, free bits + KL balancing, percentile return normalization, symexp two-hot loss) enable stable learning across over 150 tasks spanning 8 domains with continuous/discrete actions and visual/proprioceptive inputs. Dreamer matches or exceeds tuned expert algorithms in each domain and outperforms PPO everywhere. Applied out of the box, it is the first algorithm to collect diamonds in Minecraft from scratch without human data or curricula."},{"id":"arxiv:2309.05767","kind":"paper","n":203,"label":"Natural Language Supervision for General-Purpose Audio Representations (MS-CLAP 2023)","authors":"Benjamin Elizalde et al.","year":2023,"url":"https://arxiv.org/abs/2309.05767","region":"t11","hemi":"llm","lobe":"occipital","color":"#81b067","indegree":4,"alsoIn":[],"x":0.8724,"y":0.247,"text":"The paper addresses the gap between zero-shot audio-language models and task-specific models by proposing a Contrastive Language-Audio Pretraining (CLAP) model with two novel encoders: an audio encoder (HTSAT-22) pretrained on 22 audio tasks instead of just sound event classification, and an autoregressive decoder-only GPT2 text encoder adapted for sentence-level representation via a special <EOT> token. The model is pretrained on 4.6M audio-text pairs and evaluated on 26 downstream tasks\u2014the largest evaluation in the literature at the time\u2014achieving state-of-the-art on most tasks across sound events, music, speech emotion, and acoustic scenes."},{"id":"arxiv:2602.03916","kind":"paper","n":204,"label":"SpatiaLab: Can Vision-Language Models Perform Spatial Reasoning in the Wild?","authors":"SpatiaLab authors","year":2026,"url":"https://arxiv.org/abs/2602.03916","region":"t12","hemi":"llm","lobe":"frontal","color":"#5e9a4f","indegree":5,"alsoIn":[],"x":0.6106,"y":0.1555,"text":"The paper introduces SpatiaLab, a benchmark of 1,400 real-world visual question-answer pairs spanning six spatial-reasoning categories (Relative Positioning, Depth & Occlusion, Orientation, Size & Scale, Spatial Navigation, 3D Geometry) and 30 subcategories, supporting both multiple-choice (MCQ) and open-ended evaluation. Evaluating 25+ VLMs (open/closed, reasoning-focused, and spatial specialists) against human baselines reveals a large gap: the best MCQ model (InternVL3.5-72B) reaches 54.93% vs 87.57% for humans, and the best open-ended model (GPT-5-mini) reaches only 40.93% vs 64.93% for humans. Error analysis shows systematic failures in occlusion, depth, navigation, and 3D geometry, and intervention studies (SFT, CoT, CoT+self-reflection, and a multi-agent system SpatioXolver) yield only narrow, format-dependent gains rather than robust spatial competence."},{"id":"arxiv:2212.12794","kind":"paper","n":205,"label":"GraphCast: Learning skillful medium-range global weather forecasting","authors":"Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, et al.","year":2022,"url":"https://arxiv.org/abs/2212.12794","region":"t34","hemi":"llm","lobe":"temporal","color":"#9a4f93","indegree":5,"alsoIn":[],"x":0.6181,"y":0.3592,"text":"Medium-range weather forecasting is dominated by compute-intensive numerical weather prediction (NWP) that cannot learn from historical data archives. The paper introduces GraphCast, a 36.7M-parameter graph neural network in an encode-process-decode configuration with an internal multi-mesh representation, trained autoregressively on 39 years of ECMWF ERA5 reanalysis to predict 227 weather variables on a 0.25\u00b0 (721\u00d71440) global grid 6 hours ahead and rolled out to 10 days. GraphCast produces a 10-day forecast in under 60 seconds on a single Cloud TPU v4 and significantly outperforms the operational deterministic HRES system on 89.9% of 1380 verification targets. Its forecasts also better support severe-event prediction (tropical cyclone tracks, atmospheric rivers, extreme heat) despite not being trained for those tasks."},{"id":"arxiv:2309.08600","kind":"paper","n":206,"label":"Sparse Autoencoders Find Highly Interpretable Features in Language Models","authors":"Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, Lee Sharkey","year":2023,"url":"https://arxiv.org/abs/2309.08600","region":"t16","hemi":"llm","lobe":"frontal","color":"#67b084","indegree":4,"alsoIn":[],"x":0.66,"y":0.4016,"text":"The paper addresses polysemanticity in language models, where neurons activate in multiple unrelated contexts, making interpretability difficult. The authors train sparse autoencoders on internal activations (residual stream, MLP, attention) of Pythia-70M and Pythia-410M to learn overcomplete dictionaries of sparsely-activating features, hypothesized to recover features from superposition. They show these dictionary features are more interpretable than PCA, ICA, random directions, and the default basis via automated interpretability scoring, and demonstrate finer-grained causal localization on the IOI task via activation patching."},{"id":"arxiv:2502.21321","kind":"paper","n":207,"label":"LLM Post-Training: A Deep Dive into Reasoning Large Language Models","authors":"","year":null,"url":"https://arxiv.org/abs/2502.21321","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":15,"alsoIn":[],"x":0.8036,"y":0.8091,"text":"This survey systematically reviews post-training methods that refine pretrained LLMs into reasoning models, organizing the field into three interconnected pillars: supervised/parameter-efficient fine-tuning, reinforcement learning (reward modeling plus RLHF/RLAIF and policy-optimization algorithms PPO, TRPO, DPO, GRPO, ORPO, OREO), and test-time scaling (CoT, ToT, GoT, beam/best-of-N search, self-consistency, MCTS, compute-optimal scaling). It provides a taxonomy, the underlying mathematical formulations, a catalog of RL-enhanced LLMs (Table 1), a tooling/framework overview (Table 2), and a benchmark/dataset survey (Table 3). The paper's contribution is consolidation and structuring rather than new experiments, and it foregrounds open challenges such as catastrophic forgetting, reward hacking, and inference-time compute trade-offs."},{"id":"thesequence.substack.com:0f25bda15c4b","kind":"paper","n":208,"label":"The Sequence Knowledge #829: World Models and Physical AI","authors":"Jesus Rodriguez","year":2026,"url":"https://thesequence.substack.com/p/the-sequence-knowledge-829-world","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":6,"alsoIn":[],"x":0.4208,"y":0.9271,"text":"A paywalled newsletter post (TheSequence Knowledge #829) previewing a discussion of world models for physical AI, with promised coverage of NVIDIA Cosmos and World Labs' Marble. It frames a contrast between conventional world models that do 'temporal prediction' of pixels (hallucinating the next video frame) and Marble, positioned as a 'Large World Model' for reconstructing, generating, and simulating persistent 3D environments. The substantive body \u2014 the 'Lifting 2D into 4D' architecture section and the Cosmos deep-dive \u2014 is behind a paywall and not available, so no methods, experiments, or quantitative results are present."},{"id":"arxiv:2312.09257","kind":"paper","n":209,"label":"Brain-Inspired Machine Intelligence: A Survey of Neurobiologically-Plausible Credit Assignment","authors":"Alexander G. Ororbia","year":2023,"url":"https://arxiv.org/abs/2312.09257","region":"t5","hemi":"llm","lobe":"parietal","color":"#b09e67","indegree":2,"alsoIn":[],"x":0.5762,"y":0.6253,"text":"This survey reviews biologically-inspired credit assignment algorithms for training artificial neural networks without backpropagation, organizing them into a taxonomy of six families based on the question of where learning signals originate and how they are produced. The six families are: implicit signals (Hebbian), global explicit signals (feedback alignment, neuromodulation), non-synergistic local explicit signals (synthetic local updates, local signalers, recirculation), and three synergistic local explicit signal sub-families (discrepancy reduction, energy-based, forward-only). Through synthesis tables comparing each family against backprop's core problems and generalization properties, the survey finds that no single algorithm resolves all issues, but the forward-only family collectively meets all eleven evaluation criteria, and all families together cover all criteria."},{"id":"reply.com:5748177416db","kind":"paper","n":210,"label":"World Models: the operating system for spatial intelligence","authors":"Reply","year":2025,"url":"https://www.reply.com/en/artificial-intelligence/world-models-the-operating-system-for-spatial-intelligence","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":6,"alsoIn":[],"x":0.2807,"y":0.9121,"text":"This industry overview (by Reply) frames world models as the foundational 'operating system' for spatial intelligence \u2014 the ability to reason about space, motion, and physical interaction that current LLMs lack. It surveys recent world-model systems (Marble, SAM 3D, Genie 3, HunyuanWorld-Mirror, Cosmos 2.5, SIMA 2, GWM-1) and then reports a focused in-house experiment with Meta's V-JEPA 2, a predictive non-generative video architecture. The experiment confirms V-JEPA 2's strength at modeling temporal dynamics and motion, positioning world models as the bridge between perception, reasoning, and action in embodied AI and robotics."},{"id":"aclanthology.org:8f7acb76ba1d","kind":"paper","n":211,"label":"Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data","authors":"Emily M. Bender, Alexander Koller","year":2020,"url":"https://aclanthology.org/2020.acl-main.463/","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":0,"alsoIn":[],"x":0.7348,"y":0.6627,"text":"This position paper argues that language models trained purely on linguistic form (text) have, a priori, no way to learn meaning, because meaning requires a relation between form and something external to language (communicative intent, the world). Bender and Koller introduce thought experiments \u2014 most notably the 'octopus test' \u2014 to show that a learner exposed only to the statistical distribution of form can mimic fluent text without acquiring the grounded understanding that NLU implies. They push back on hype framing large neural LMs as 'understanding' language, and propose terminological and methodological hygiene to keep NLU research scientifically grounded. The contribution is conceptual: clarifying the form/meaning distinction to guide better science around natural language understanding."},{"id":"arxiv:2112.10752","kind":"paper","n":212,"label":"High-Resolution Image Synthesis with Latent Diffusion Models","authors":"Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, Bj\u00f6rn Ommer","year":2022,"url":"https://arxiv.org/abs/2112.10752","region":"t32","hemi":"wm","lobe":"frontal","color":"#8b4f9a","indegree":7,"alsoIn":[],"x":0.4492,"y":0.1668,"text":"Diffusion models produce state-of-the-art image synthesis but operate directly in high-dimensional pixel space, making training (hundreds of GPU days) and sampling expensive. This paper proposes Latent Diffusion Models (LDMs): a pretrained autoencoder first compresses images into a perceptually-equivalent lower-dimensional latent space, and the diffusion model is then trained in that latent space, with a cross-attention mechanism added to the UNet for general conditioning (text, layouts, semantic maps, class labels). LDMs set new state-of-the-art FID scores on CelebA-HQ (5.11) and image inpainting while matching or beating prior diffusion/AR/GAN methods on class-conditional ImageNet, text-to-image, super-resolution and unconditional generation, all at substantially reduced compute and parameter cost."},{"id":"arxiv:2509.14252","kind":"paper","n":213,"label":"LLM-JEPA: Large Language Models Meet Joint Embedding Predictive Architectures","authors":"Hai Huang, Yann LeCun, Randall Balestriero","year":2025,"url":"https://arxiv.org/abs/2509.14252","region":"t1","hemi":"wm","lobe":"parietal","color":"#b07267","indegree":2,"alsoIn":[],"x":0.2108,"y":0.3795,"text":"The paper introduces LLM-JEPA, the first adaptation of Joint Embedding Predictive Architectures (JEPAs) to language model training, combining a standard next-token prediction loss with an embedding-space JEPA objective that aligns representations of two views (e.g., natural language text and code) of the same underlying knowledge. The method is applied to both fine-tuning and pretraining across multiple model families and datasets, consistently improving accuracy while resisting overfitting. Results show statistically significant gains across Llama3, Gemma2, OpenELM, OLMo, Qwen3, and DeepSeek-R1-Distill models on tasks including NL-to-regex, NL-to-SQL, GSM8K, NQ-Open, and HellaSwag."},{"id":"arxiv:1903.00374","kind":"paper","n":214,"label":"Model-Based Reinforcement Learning for Atari","authors":"Lukasz Kaiser, Mohammad Babaeizadeh, Piotr Milos, Blazej Osinski, Roy H. Campbel","year":2019,"url":"https://arxiv.org/abs/1903.00374","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":5,"alsoIn":[],"x":0.1719,"y":0.2628,"text":"Addresses the poor sample efficiency of model-free RL on Atari, which needs tens-to-hundreds of millions of steps where humans learn in minutes. Proposes SimPLe, a model-based deep RL algorithm that learns an action-conditioned video prediction world model (including a novel stochastic model with discrete latent variables) from real interactions, then trains a PPO policy entirely inside the learned simulator, iterating data-collection / model-training / policy-training 15 times. Evaluated in a 100K-interaction (\u2248400K frames, ~2 hours play) low-data regime on 26 Atari games, SimPLe is substantially more sample-efficient than highly-tuned Rainbow and PPO, beating them on most games and by over an order of magnitude on Freeway. The novel stochastic-discrete world model gives the best results among the architectures tried, though final absolute scores remain below the best fully-trained model-free methods."},{"id":"arxiv:2204.01691","kind":"paper","n":215,"label":"Do As I Can, Not As I Say: Grounding Language in Robotic Affordances","authors":"Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron Dav","year":2022,"url":"https://arxiv.org/abs/2204.01691","region":"t0","hemi":"wm","lobe":"frontal","color":"#9a4f4f","indegree":0,"alsoIn":[],"x":0.1492,"y":0.733,"text":"SayCan addresses how an embodied robot can exploit the semantic knowledge in large language models, which lack physical grounding and may propose actions that are infeasible for a given robot or scene. The method combines an LLM ('Say'), which scores how useful each candidate low-level skill is for a high-level natural-language instruction, with learned value functions ('Can'), which give the affordance probability that each skill can succeed from the current state; multiplying the two selects the skill that is both useful and feasible, then iterates until 'done'. Evaluated on 101 real-world instructions across 7 families on a mobile manipulator in office kitchens, PaLM-SayCan reaches 84% planning and 74% execution success in the mock kitchen, with affordance grounding nearly doubling performance over non-grounded baselines. The paper also shows robot performance scales with the underlying LLM (PaLM 540B beats FLAN), and demonstrates chain-of-thought, multilingual, and easily-added new skills (drawer manipulation)."},{"id":"arxiv:2404.08471","kind":"paper","n":216,"label":"Revisiting Feature Prediction for Learning Visual Representations from Video (V-JEPA)","authors":"Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann Le","year":2024,"url":"https://arxiv.org/abs/2404.08471","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":9,"alsoIn":[],"x":0.2859,"y":0.4674,"text":"The paper investigates whether feature prediction alone can serve as a stand-alone self-supervised objective for learning visual representations from video, introducing V-JEPA \u2014 a family of ViT models trained to predict the representations (not pixels) of masked spatio-temporal regions from a visible context, using an EMA target encoder, stop-gradient, predictor, and an L1 loss, with no pretrained image encoders, text, negatives, or pixel reconstruction. Models are pretrained on VideoMix2M (~2M public videos) and evaluated frozen (attentive probing) and fine-tuned on video and image tasks. The largest model (ViT-H/16 384) reaches 81.9% on Kinetics-400, 72.2% on Something-Something-v2, and 77.9% on ImageNet-1K, beating prior video models on motion tasks and narrowing the gap to image models on appearance tasks. Feature prediction is shown to be more efficient (fewer pretraining samples, ~2\u00d7 faster) and more label-efficient than pixel-reconstruction approaches."},{"id":"arxiv:2509.12387","kind":"paper","n":217,"label":"Causal-Symbolic Meta-Learning (CSML): Inducing Causal World Models for Few-Shot Generalization","authors":"Anonymous et al.","year":2025,"url":"https://arxiv.org/abs/2509.12387","region":"t6","hemi":"llm","lobe":"temporal","color":"#b0a967","indegree":4,"alsoIn":[],"x":0.6685,"y":0.6928,"text":"CSML addresses the problem that deep learning models rely on spurious correlations, leading to poor few-shot generalization. The framework combines a ViT-based perception module that produces disentangled symbolic latents, a differentiable causal induction module (adapted from NOTEARS) that discovers a shared DAG across tasks, and a GCN-based reasoning module that performs message passing on the induced causal graph. Evaluated on the newly introduced CausalWorld 2D physics benchmark, CSML dramatically outperforms MAML, ProtoNets, and a fixed-graph neuro-symbolic baseline, especially on 0-shot intervention and counterfactual tasks."},{"id":"arxiv:2508.12027","kind":"paper","n":218,"label":"Active inference for action-unaware agents","authors":"Filippo Torresan, Keisuke Suzuki, Ryota Kanai","year":2025,"url":"https://arxiv.org/abs/2508.12027","region":"t14","hemi":"llm","lobe":"temporal","color":"#67b06e","indegree":3,"alsoIn":[],"x":0.5732,"y":0.8982,"text":"The paper compares two formulations of discrete-time (POMDP) active inference agents that differ in whether an agent has access to a copy of its own past actions: action-aware agents (which know their executed actions, akin to an efference copy / control-as-inference) and action-unaware agents (which must infer the policy/actions they performed from past observations, following classical active-inference and referent-control traditions). The authors provide a Python implementation and run simulations of both agent types in two navigation tasks \u2014 a 4-step T-maze and a 3\u00d73 grid world \u2014 analyzing free energy, expected free energy, policy-conditioned free energies, and policy probabilities. They find that action-unaware agents can match the goal-reaching performance of action-aware agents despite being at a severe informational and computational disadvantage, but at a much higher cost: perceptual inference is O(n) for action-aware vs O(nm) for action-unaware agents (n past states, m policies), causing a combinatorial explosion that limits scalability."},{"id":"arxiv:2208.06726","kind":"paper","n":219,"label":"Predicting the cascading dynamics in complex networks via the bimodal failure size distribution","authors":"","year":null,"url":"https://arxiv.org/abs/2208.06726","region":"t36","hemi":"llm","lobe":"frontal","color":"#b06793","indegree":7,"alsoIn":[],"x":0.6301,"y":0.7414,"text":"The paper investigates why cascade sizes in complex networks follow a bimodal distribution (cascades are either very small or catastrophically large) under the Motter-Lai overload model, and how to predict which a single-node failure will trigger. The authors show the bimodal pattern is ubiquitous across synthetic (ER, SW, BA) and real networks, and find that large (right-peak) cascades arise either from high-load nodes failing at the first step (high MLF) or from many sequential cascade rounds (high T), whereas the size of First Failures itself does not predict the outcome. They propose a Hybrid Load Metric (HLM) combining the initial node's load with the maximal load of First Failures, and validate it via AUC. HLM outperforms betweenness/load and other centrality metrics at classifying cascades into the left vs right peak across all tested networks."},{"id":"arxiv:2303.08302","kind":"paper","n":220,"label":"ZeroQuant-V2: Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation","authors":"Zhewei Yao, Xiaoxia Wu, Cheng Li, et al.","year":2023,"url":"https://arxiv.org/abs/2303.08302","region":"t37","hemi":"llm","lobe":"parietal","color":"#b06788","indegree":4,"alsoIn":[],"x":0.8769,"y":0.3643,"text":"The paper conducts a systematic empirical study of post-training quantization (PTQ) for LLMs across weight-only, activation-only, and weight-and-activation schemes, using RTN, GPTQ, and ZeroQuant (Global/Local) variants on OPT (125M\u201366B) and BLOOM (560M\u2013176B). It finds activation quantization is more sensitive than weight quantization, smaller models tolerate activation quantization better, and no existing method fully recovers FP16 quality at INT4-weight or W4A8. It proposes Low-Rank Compensation (LoRC), which factorizes the quantization error matrix into two low-rank INT8 matrices to recover model quality with minimal (<1.6%) size increase, and shows LoRC combined with fine-grained quantization sets a new size-quality Pareto frontier."},{"id":"arxiv:2202.03771","kind":"paper","n":221,"label":"Energy Management Based on Multi-Agent Deep Reinforcement Learning for A Multi-Energy Industrial Park","authors":"","year":null,"url":"https://arxiv.org/abs/2202.03771","region":"t4","hemi":"wm","lobe":"frontal","color":"#9a7c4f","indegree":3,"alsoIn":[],"x":0.2937,"y":0.1631,"text":"The paper tackles multi-energy (electricity/heat/gas) management in an industrial park where coupled energy sources, uncertain renewables, and stochastic demand make centralized scheduling computationally heavy. It formulates the problem as a partially-observable MDP and proposes a multi-agent deep reinforcement learning framework with centralized training and decentralized execution, combining a counterfactual advantage baseline, discrete soft actor-critic, a Lagrange-multiplier reward for storage capacity constraints, and an attention mechanism for scalability. On real price/load/PV data, the proposed algorithm reaches a total cost of \u00a574,133, within ~3.7% of the exhaustive-method optimum (\u00a571,493) and below DDPG, MADDPG, and a restricted no-attention version. Improvements over baselines grow with agent count, showing better scalability."},{"id":"arxiv:2301.04104","kind":"paper","n":222,"label":"Mastering Diverse Domains through World Models","authors":"Hafner et al.","year":2023,"url":"https://arxiv.org/abs/2301.04104","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":4,"alsoIn":[],"x":0.1304,"y":0.3757,"text":"The paper presents DreamerV3, a model-based reinforcement learning algorithm that learns a world model (RSSM) and trains an actor-critic purely on imagined trajectories, using a single fixed hyperparameter configuration across all domains. Robustness techniques\u2014symlog transforms, symexp twohot loss, KL balancing with free bits, percentile return normalization, and 1% unimix categoricals\u2014enable stable learning across over 150 tasks spanning 8 benchmarks with continuous/discrete actions, visual/vector inputs, and dense/sparse rewards. Dreamer outperforms tuned specialist algorithms (MuZero, Rainbow, IQN, PPG, IMPALA, DrQ-v2, D4PG) on each benchmark and beats a high-quality PPO baseline everywhere. Applied out of the box, it is the first algorithm to collect diamonds in Minecraft from scratch from sparse rewards without human data or curricula."},{"id":"arxiv:2510.13161","kind":"paper","n":223,"label":"Mirror Speculative Decoding: Breaking the Serial Barrier in LLM Inference","authors":"","year":null,"url":"https://arxiv.org/abs/2510.13161","region":"t9","hemi":"llm","lobe":"parietal","color":"#97b067","indegree":2,"alsoIn":[],"x":0.8223,"y":0.7782,"text":"LLM speculative decoding is bottlenecked by serial autoregressive draft generation, where larger drafts raise acceptance but add latency overhead, capping speedup. The paper introduces Mirror Speculative Decoding (Mirror-SD), a systems-algorithm co-design that launches branch-complete draft rollouts from the target's early-exit signals in parallel with the target's suffix computation, maps the draft onto NPUs and target onto GPUs across heterogeneous accelerators, and pairs this with speculative streaming (SS) so the draft emits multiple tokens per step. On SpecBench with 14B-66B server-scale models, Mirror-SD delivers 2.8x-5.8x wall-time speedups, a ~30% average relative improvement over the strongest baseline EAGLE3, while remaining lossless (identical outputs to the target)."},{"id":"arxiv:2401.03804","kind":"paper","n":224,"label":"TeleChat Technical Report","authors":"Zihan Wang et al. (China Telecom)","year":2024,"url":"https://arxiv.org/abs/2401.03804","region":"t10","hemi":"llm","lobe":"temporal","color":"#8cb067","indegree":1,"alsoIn":[],"x":0.7966,"y":0.3915,"text":"This technical report introduces TeleChat, a family of bilingual (Chinese/English) autoregressive transformer LLMs at 3B, 7B, and 12B parameters, pretrained on 0.8T\u20131.2T tokens and aligned via supervised fine-tuning, NEFTune noisy-embedding tuning, multi-stage long-context training, and PPO-based reinforcement learning. The report emphasizes transparency about its data-cleaning pipeline (hierarchical dedup, KenLM perplexity filtering) and SFT methodology, and extends the context window beyond 96k tokens using NTK-aware interpolation plus LogN-Scaling. TeleChat-7B-chat achieves competitive-to-superior scores against similar-size open models on exam, understanding, reasoning, and coding benchmarks, and the authors show a knowledge-graph retrieval method that raises factual-QA accuracy from 0.19 to 0.69. The 7B and 12B fine-tuned checkpoints, code, and 1TB of pretraining data are released."},{"id":"arxiv:2411.06284","kind":"paper","n":225,"label":"A Comprehensive Survey and Guide to Multimodal Large Language Models in Vision-Language Tasks","authors":"","year":null,"url":"https://arxiv.org/abs/2411.06284","region":"t30","hemi":"llm","lobe":"temporal","color":"#8c67b0","indegree":6,"alsoIn":[],"x":0.7853,"y":0.2307,"text":"A broad survey and practitioner guide to Multimodal Large Language Models (MLLMs) for vision-language tasks, tracing the lineage from traditional NLP and LLMs (Transformers, BERT, GPT) to multimodal architectures built from visual encoders, language decoders, multimodal fusion modules, and cross-attention layers. It systematizes training methodology (contrastive pre-training \u00e0 la CLIP/ALIGN, masked language/image modeling, VQA/VLP pre-training, fine-tuning, few-/zero-shot and instruction tuning) and catalogs applications across image captioning, VQA, storytelling, content creation, cross-modal retrieval, and accessibility. It presents case studies of dozens of deployed systems (image generation, code generation, RAG, robotics/embodied AI, video/audio) and closes with challenges in scalability, robustness, interpretability, evaluation, and ethics. As a survey it contributes a taxonomy and reference rather than new experimental results."},{"id":"arxiv:2405.15383","kind":"paper","n":226,"label":"Generating Code World Models with Large Language Models Guided by Monte Carlo Tree Search","authors":"Dainese et al.","year":2024,"url":"https://arxiv.org/abs/2405.15383","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":2,"alsoIn":[],"x":0.8006,"y":0.7931,"text":"The paper introduces Code World Models (CWMs) \u2014 environment dynamics and reward functions written as executable Python code by an LLM for model-based RL \u2014 and GIF-MCTS (Generate, Improve, Fix with Monte Carlo Tree Search), a code-generation strategy that iteratively builds and self-debugs programs using accuracy feedback from environment trajectories as unit tests. It also releases the Code World Models Benchmark (CWMB) of 18 discrete and continuous control RL environments paired with language descriptions and curated offline trajectories. GIF-MCTS beats WorldCoder and other baselines on CWMB, the APPS Competition split, and RTFM, and the synthesized CWMs enable model-based planning that is four-to-seven orders of magnitude faster at inference than querying an LLM as the world model."},{"id":"arxiv:2506.06725","kind":"paper","n":227,"label":"WorldLLM: Improving LLMs' world modeling using curiosity-driven theory-making","authors":"Lehman, Perez et al.","year":2025,"url":"https://arxiv.org/abs/2506.06725","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":2,"alsoIn":[],"x":0.6974,"y":0.8275,"text":"WorldLLM tackles the problem that LLMs hold broad world knowledge but make poor precise predictions of dynamics in structured, domain-specific environments. The framework improves an LLM-based forward model (the 'Statistician', P(s'|s,a,H)) by conditioning its predictions on natural-language hypotheses, which a second LLM (the 'Scientist') iteratively refines via Metropolis Bayesian inference, while a curiosity-driven RL agent (the 'Experimenter') collects low-likelihood transitions as new evidence. In the Playground-Text object-combination game, the loop measurably raises predictive log-likelihood and top-3 next-state accuracy over a no-hypothesis baseline and produces human-interpretable theories of environment dynamics. Compared to fine-tuning, the prompted-hypothesis approach avoids gradient updates and is interpretable, but both methods generalize poorly to syntax changes."},{"id":"arxiv:2406.09455","kind":"paper","n":228,"label":"Pandora: Towards General World Model with Natural Language Actions and Video States","authors":"Xiang, Liu et al. (Maitrix.org)","year":2024,"url":"https://arxiv.org/abs/2406.09455","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":5,"alsoIn":[],"x":0.3092,"y":0.8262,"text":"Pandora is a hybrid autoregressive-diffusion world model that simulates future world states as videos while accepting free-text natural-language actions at any point during generation, enabling on-the-fly interactive control rather than text prompts only at the start. It avoids training from scratch by stitching a pretrained LLM (Vicuna-7B-v1.5 via Chat-UniVi) to a pretrained text-to-video diffusion model (DynamiCrafter) using a vision encoder and Q-Former adapters, trained in two stages: embedding-alignment pretraining (L2 loss on WebVid-10M, backbones frozen) and instruction tuning on a curated 1.2M video-action dataset. Results are qualitative across diverse domains (indoor/outdoor, human/robot, 2D/3D games, driving, physical phenomena); the authors report longer-video generation and emergent cross-domain controllability with more training compute, and explicitly defer quantitative evaluation to future work."},{"id":"lumalabs.ai:cf7b765e8f15","kind":"paper","n":229,"label":"Breaking the Algorithmic Ceiling in Pre-Training with Inductive Moment Matching","authors":"Linqi Zhou, Stefano Ermon, Jiaming Song (Luma AI)","year":2025,"url":"https://lumalabs.ai/news/inductive-moment-matching","region":"t3","hemi":"llm","lobe":"occipital","color":"#9a714f","indegree":0,"alsoIn":[],"x":0.7243,"y":0.4084,"text":"The paper argues that generative pre-training's apparent limits stem from algorithmic stagnation (only autoregressive and diffusion paradigms since ~2020) rather than data scarcity, and introduces Inductive Moment Matching (IMM), a new pre-training method. IMM augments the network to condition on the target timestep (not just the current one) alongside a maximum-mean-discrepancy moment-matching objective, making each inference iteration flexible enough for high-quality few-step generation. On ImageNet 256x256 it reaches 1.99 FID, beating diffusion (2.27) and Flow Matching (2.15) with 30x fewer sampling steps, and hits a 2-step FID of 1.98 on CIFAR-10 trained from scratch. Unlike consistency models, which train unstably and collapse, IMM uses a single objective that remains stable across hyperparameters and architectures."},{"id":"nature.com:386c3df42560","kind":"paper","n":230,"label":"Probabilistic weather forecasting with machine learning (GenCast)","authors":"Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, et al.","year":2024,"url":"https://www.nature.com/articles/s41586-024-08252-9","region":"t34","hemi":"wm","lobe":"temporal","color":"#9a4f93","indegree":7,"alsoIn":[],"x":0.2496,"y":0.1325,"text":"GenCast is a probabilistic medium-range weather forecasting model implemented as a conditional diffusion model that generates ensembles of 15-day global forecasts at 0.25\u00b0 resolution and 12-h steps for 80+ surface/atmospheric variables, trained on 40 years (1979-2018) of ERA5 reanalysis. It models the conditional distribution of the next atmospheric state given the two previous states via iterative denoising on a refined icosahedral mesh processed by a graph transformer, producing sharp, spectrally realistic ensemble members rather than blurry means. Evaluated on 2019, it outperforms ECMWF's ENS \u2014 the top operational NWP ensemble \u2014 on 97.2% of 1,320 CRPS targets, with better extreme-weather, tropical-cyclone-track and wind-power forecasts, while generating a forecast in ~8 minutes on a Cloud TPUv5. It is, to the authors' knowledge, the first ML weather model to significantly beat ENS on probabilistic skill."},{"id":"arxiv:2006.04182","kind":"paper","n":231,"label":"Predictive Coding Approximates Backprop along Arbitrary Computation Graphs","authors":"Beren Millidge, Alexander Tschantz, Christopher L. Buckley","year":2020,"url":"https://arxiv.org/abs/2006.04182","region":"t20","hemi":"llm","lobe":"frontal","color":"#67b0b0","indegree":8,"alsoIn":[],"x":0.6286,"y":0.6491,"text":"The paper proves that predictive coding \u2014 a biologically-plausible cortical process theory using only local, Hebbian updates \u2014 converges asymptotically to the exact backpropagation gradients on arbitrary computation graphs, not just layered MLPs. The authors derive predictive coding as variational inference under a hierarchical Gaussian generative model and show that, at the equilibrium of its free-energy dynamics, prediction-error units exactly equal the backpropagated gradients via the same recursive chain-rule structure. They give a recipe to translate any differentiable architecture into a predictive-coding equivalent and build PC versions of CNNs, RNNs, and LSTMs. These models match backprop performance on tasks harder than MNIST while using only local and (for parameter-linear layers) Hebbian plasticity."},{"id":"arxiv:2103.04723","kind":"paper","n":232,"label":"Optimal Scheduling of Integrated Demand Response-Enabled Integrated Energy Systems with Uncertain Renewable Generations: A Stackelberg Game Approach","authors":"","year":null,"url":"https://arxiv.org/abs/2103.04723","region":"t4","hemi":"wm","lobe":"frontal","color":"#9a7c4f","indegree":5,"alsoIn":[],"x":0.2153,"y":0.2348,"text":"The paper proposes a Stackelberg game-based optimization framework for scheduling an integrated demand response (IDR)-enabled integrated energy system (IES) with uncertain wind/PV generation, where an integrated energy operator (IEO) leads by setting real-time electricity/heat prices to maximize profit and users follow by adjusting consumption to minimize cost. Renewable uncertainty is handled via a chance-constrained spinning reserve converted to a deterministic form using sequence operation theory (SOT), a district heating network (DHN) model with time delay and thermal attenuation is built, and user thermal comfort is captured by a predicted mean vote (PMV) index. The bi-level game is recast into a mixed-integer quadratic program via KKT conditions plus the Big-M linearization and solved with CPLEX. Tests on a modified IEEE 30-bus + two 6-bus district heating systems and a real IES in Jilin, China show the method reaches a Stackelberg equilibrium and fully absorbs available renewables under the joint-optimization mode."},{"id":"arxiv:2305.18153","kind":"paper","n":233,"label":"Do Large Language Models Know What They Don't Know?","authors":"Yin et al. (Zhangyue Yin, Qiushi Sun, Qipeng Guo, Jiawen Wu, Xipeng Qiu, Xuanjin","year":2023,"url":"https://arxiv.org/abs/2305.18153","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":2,"alsoIn":[],"x":0.8275,"y":0.7631,"text":"The paper investigates whether LLMs possess 'self-knowledge' \u2014 the ability to recognize unanswerable or unknowable questions and express uncertainty rather than fabricate answers. The authors introduce SelfAware, a dataset of 1,032 unanswerable questions across five categories plus 2,337 answerable counterparts, and an automated F1-based evaluation that detects uncertainty by measuring SimCSE similarity between model outputs and 16 curated reference 'uncertainty' sentences (threshold 0.75). Evaluating 20 LLMs (GPT-3, InstructGPT, LLaMA, Alpaca, Vicuna, GPT-4), they find self-knowledge scales with model size and is boosted by instruction tuning and in-context learning. The best model, GPT-4, reaches 75.47% F1 versus a human benchmark of 84.93%, revealing a substantial gap."},{"id":"arxiv:2403.00504","kind":"paper","n":234,"label":"Learning and Leveraging World Models in Visual Representation Learning","authors":"Quentin Garrido, Mahmoud Assran, Nicolas Ballas, Adrien Bardes, Laurent Najman, ","year":2024,"url":"https://arxiv.org/abs/2403.00504","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":3,"alsoIn":[],"x":0.2522,"y":0.4246,"text":"The paper introduces Image World Models (IWM), a JEPA-based self-supervised approach that goes beyond masked image modeling by training a latent predictor (world model) conditioned on photometric and destructive augmentation parameters to predict their effect in latent space. The authors identify three ingredients for a capable (equivariant) world model \u2014 conditioning on transformations, sufficient transformation difficulty, and predictor capacity \u2014 and show the learned predictor can be reused as an efficient evaluation head via 'predictor finetuning'. Fine-tuning the equivariant predictor on a frozen encoder matches or beats encoder finetuning at a fraction of the cost, and the world model's capacity controls whether representations are invariant (best for linear probing) or equivariant (best for finetuning/OOD)."},{"id":"huggingface.co:9f24e732f5fe","kind":"paper","n":235,"label":"Generation Models Know Space: Unleashing Implicit 3D Priors for Scene Understanding","authors":"authors","year":2026,"url":"https://huggingface.co/papers/2603.19235","region":"t15","hemi":"wm","lobe":"occipital","color":"#4f9a62","indegree":4,"alsoIn":[],"x":0.4051,"y":0.6755,"text":"MLLMs suffer from 'spatial blindness'\u2014weak fine-grained geometric reasoning and physical dynamics\u2014and prior fixes rely on explicit 3D modalities or geometric scaffolding that face data scarcity and generalization limits. The paper introduces VEGA-3D (Video Extracted Generative Awareness), a plug-and-play framework that repurposes a frozen pre-trained video diffusion model as a Latent World Simulator, extracting spatiotemporal features from intermediate denoising noise levels and fusing them into MLLM semantic representations via a token-level adaptive gated fusion mechanism. This injects dense geometric/physical priors without explicit 3D supervision. The authors report that VEGA-3D outperforms state-of-the-art baselines across 3D scene understanding, spatial reasoning, and embodied manipulation benchmarks."},{"id":"arxiv:2305.07017","kind":"paper","n":236,"label":"An Inverse Scaling Law for CLIP Training","authors":"Xianhang Li, Zeyu Wang, Cihang Xie","year":2023,"url":"https://arxiv.org/abs/2305.07017","region":"t23","hemi":"llm","lobe":"occipital","color":"#4f799a","indegree":2,"alsoIn":[],"x":0.6794,"y":0.1261,"text":"CLIP training is prohibitively expensive, limiting exploration to well-resourced labs. This paper discovers an inverse scaling law: larger image/text encoders can be trained with shorter image/text token sequences while retaining competitive performance, and the token-reduction strategy that best preserves semantic information (image resizing, syntax masking) yields the best scaling. Leveraging this, the authors introduce CLIPA, which trains high-accuracy CLIP models cheaply \u2014 e.g., 69.3% ImageNet-1k zero-shot in ~4 days on 8 A100s, and a record 83.0% zero-shot accuracy with G/14 at ~33\u00d7 less compute than OpenCLIP-G/14."},{"id":"doi.org:0a734c9c22c9","kind":"paper","n":237,"label":"Human-level control through deep reinforcement learning","authors":"Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Ma","year":2015,"url":"https://doi.org/10.1038/nature14236","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":3,"alsoIn":[],"x":0.2419,"y":0.1444,"text":"The paper addresses how an agent can learn control policies directly from high-dimensional sensory input rather than hand-engineered features. It introduces the deep Q-network (DQN), which combines Q-learning with a deep convolutional neural network, plus experience replay and a separate target network to stabilize training, learning end-to-end from raw Atari 2600 pixels and the game score. Using one identical algorithm, architecture, and hyperparameter set across 49 games, DQN surpassed all prior RL algorithms and reached a level comparable to a professional human games tester. This is presented as the first single artificial agent able to learn to excel at a diverse array of challenging tasks from sensory input."},{"id":"arxiv:2406.11944","kind":"paper","n":238,"label":"Transcoders Find Interpretable LLM Feature Circuits","authors":"Jacob Dunefsky, Philippe Chlenski, Neel Nanda","year":2024,"url":"https://arxiv.org/abs/2406.11944","region":"t16","hemi":"llm","lobe":"frontal","color":"#67b084","indegree":7,"alsoIn":[],"x":0.6989,"y":0.4599,"text":"MLP sublayers make fine-grained circuit analysis on transformer LLMs hard because interpretable SAE features are dense combinations of many nonlinear neurons, and SAE-based attributions are input-dependent so they can't describe an MLP's general behavior. This paper explores transcoders\u2014wide, sparsely-activating ReLU MLPs trained to approximate an MLP sublayer's input-output computation with an L1 sparsity penalty\u2014and introduces a weights-based circuit analysis method whose feature-pair attributions factorize into an input-dependent activation term and an input-invariant decoder\u00b7encoder dot product. Trained on GPT2-small, Pythia-410M, and Pythia-1.4B, transcoders match or beat MLP-output SAEs on sparsity, faithfulness, and human interpretability, with the gap widening on larger models. Applied to case studies including the GPT2-small 'greater-than' circuit, transcoders recover most of the model's behavior with far fewer features than neurons and support 'blind' reverse-engineering of unknown features."},{"id":"arxiv:2206.11795","kind":"paper","n":239,"label":"Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online Videos","authors":"Bowen Baker, Ilge Akkaya, Peter Zhokhov, Joost Huizinga, Jie Tang, Adrien Ecoffe","year":2022,"url":"https://arxiv.org/abs/2206.11795","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":0,"alsoIn":[],"x":0.2252,"y":0.5943,"text":"VPT (Video PreTraining) extends internet-scale pretraining to sequential decision domains via semi-supervised imitation learning: a small amount of labeled contractor data trains a non-causal inverse dynamics model (IDM) that pseudo-labels ~70k hours of unlabeled online Minecraft videos, on which a causal behavioral-cloning foundation model is trained. The 0.5B-parameter foundation model shows nontrivial zero-shot skills (chopping trees, crafting planks/tables) and can be further fine-tuned with BC and RL. RL fine-tuning produces an agent that crafts a diamond pickaxe \u2014 a task taking proficient humans >20 minutes (~24,000 actions) \u2014 the first reported computer agent to do so, while operating the native human mouse-and-keyboard interface at 20Hz. A core finding is that contractor data is far more efficiently used to train an IDM within VPT than to train a BC model directly."},{"id":"arxiv:2111.00210","kind":"paper","n":240,"label":"Mastering Atari Games with Limited Data","authors":"Weirui Ye, Shaohuai Liu, Thanard Kurutach, Pieter Abbeel, Yang Gao","year":2021,"url":"https://arxiv.org/abs/2111.00210","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":2,"alsoIn":[],"x":0.251,"y":0.1913,"text":"Addresses the sample-inefficiency of image-based RL, where consistent human-level Atari performance with limited data remained unsolved. Proposes EfficientZero, a model-based visual RL algorithm built on MuZero with three additions: a self-supervised temporal consistency loss for the environment model, end-to-end prediction of the value prefix to alleviate compounding/state-aliasing error, and model-based off-policy correction of value targets. On the Atari 100k benchmark (2 hours of gameplay) it reaches 194.3% mean and 109.0% median human-normalized score \u2014 the first super-human result at such low data \u2014 and matches state-based SAC on DMControl 100k tasks."},{"id":"arxiv:1709.02341","kind":"paper","n":241,"label":"Deep Active Inference","authors":"Kai Ueltzh\u00f6ffer","year":2017,"url":"https://arxiv.org/abs/1709.02341","region":"t14","hemi":"llm","lobe":"temporal","color":"#67b06e","indegree":6,"alsoIn":[],"x":0.5628,"y":0.8885,"text":"The paper introduces the 'Deep Active Inference' agent, which implements the free energy principle / active inference from cognitive neuroscience using deep neural networks for the generative model, an amortised inference network for the variational density, and an action function, all trained jointly. Because the agent has no access to the environment's equations of motion (or their derivatives) and the objective is non-differentiable, gradients on the variational free energy bound are estimated with evolution strategies over a population of agents. On the discrete-time mountain car problem the agent learns goal-directed behaviour by encoding the goal as a prior expectation on a latent state, while simultaneously learning a generative model of its environment that can be freely sampled or constrained-sampled (MCMC) to read out its beliefs. The agent discovers the non-trivial momentum strategy (first move away from the target) and converges within 30,000 steps in under 3.5 hours."},{"id":"arxiv:2311.16038","kind":"paper","n":242,"label":"OccWorld: Learning a 3D Occupancy World Model for Autonomous Driving","authors":"Zheng et al.","year":2023,"url":"https://arxiv.org/abs/2311.16038","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":0,"alsoIn":[],"x":0.4259,"y":0.7848,"text":"Most autonomous-driving prediction pipelines forecast only the movement of object bounding boxes, missing fine-grained 3D scene structure and requiring expensive instance/map annotations. OccWorld learns a self-supervised world model directly in 3D semantic occupancy space: a VQ-VAE scene tokenizer discretizes occupancy into high-level tokens, and a GPT-like spatial-temporal generative transformer autoregressively predicts future scene tokens and ego tokens to jointly forecast scene evolution and ego trajectory. On nuScenes/Occ3D it achieves an average IoU of 26.63 and mIoU of 17.14 for 3s-future forecasting given 2s history, and produces competitive planning (avg L2 1.17) without any instance or map supervision. It can even generate more reasonable drivable areas than ground truth, indicating scene understanding rather than memorization."},{"id":"arxiv:2412.18106","kind":"paper","n":243,"label":"Tackling the Dynamicity in a Production LLM Serving System with SOTA Optimizations via Hybrid Prefill/Decode/Verify Scheduling on Efficient Meta-kernels","authors":"","year":null,"url":"https://arxiv.org/abs/2412.18106","region":"t24","hemi":"llm","lobe":"frontal","color":"#6784b0","indegree":3,"alsoIn":[],"x":0.9068,"y":0.5577,"text":"Production LLM serving on Ascend NPUs (tile-based DSAs) suffers from severe workload variability because dynamic input/output lengths are compounded by optimizations like Automatic Prefix Caching, Speculative Decoding, and SplitFuse, which mix Prefill/Decode/Verify (P/D/V) stages with arbitrary attention shapes and matrix sizes. XY-Serve, an Ascend-native end-to-end serving system built on vLLM, introduces an abstraction layer that decomposes dynamic workloads into fixed, hardware-friendly meta-primitives: a Meta-Attention kernel that computes a unified matmul-softmax-matmul pattern via dynamic tiling and task reordering, and SmoothGEMM which uses on-chip virtual padding plus selective HBM reads/writes to run arbitrary-shaped GEMMs at fixed-tile efficiency without padding overhead. It reports up to 89% end-to-end throughput improvement over Ascend-vLLM, with GEMM kernels averaging 14.6% faster and attention kernels 21.5% faster than existing torch-npu libraries. On end-to-end MFU/MBU it performs on par with an Nvidia A800, showing up to 17% higher MBU."},{"id":"arxiv:2306.14066","kind":"paper","n":244,"label":"SEEDS: Emulation of Weather Forecast Ensembles with Diffusion Models","authors":"Lizao Li, Rob Carver, Ignacio Lopez-Gomez, Fei Sha, John Anderson","year":2023,"url":"https://arxiv.org/abs/2306.14066","region":"t34","hemi":"wm","lobe":"temporal","color":"#9a4f93","indegree":5,"alsoIn":[],"x":0.247,"y":0.1347,"text":"Operational weather centers can only afford 10-50 physics-based ensemble members per forecast cycle, too few to assess rare/extreme events. The authors propose SEEDS (Scalable Ensemble Envelope Diffusion Sampler), a conditional diffusion model with an axial Vision-Transformer score network that, given as few as K=2 seeding members from the GEFS operational ensemble, generates hundreds to tens of thousands of additional weather-like forecasts at negligible cost. The emulation model (seeds-gee) matches the full 31-member GEFS in skill metrics (RMSE, ACC, CRPS, rank histogram, Brier score), while the post-processing model (seeds-gpp), which blends in ERA5 reanalysis, debiases the forecasts and beats GEFS on reliability and near-surface temperature, especially for extreme (\u00b12\u03c3/\u00b13\u03c3) events."},{"id":"arxiv:2505.09694","kind":"paper","n":245,"label":"EWMBench: Evaluating Scene, Motion, and Semantic Quality in Embodied World Models","authors":"Li et al.","year":2025,"url":"https://arxiv.org/abs/2505.09694","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":3,"alsoIn":[],"x":0.3196,"y":0.8729,"text":"The paper introduces EWMBench, the first evaluation benchmark tailored to embodied world models (EWMs) \u2014 text-to-video/image-to-video diffusion models repurposed to generate physically actionable robot manipulation scenes from language instructions. It evaluates generated videos along three axes \u2014 visual scene consistency, motion correctness, and semantic alignment/diversity \u2014 using a curated dataset of 10 action-ordering manipulation tasks drawn from AgiBot-World plus an open-source toolkit combining a fine-tuned DINOv2, a YOLO-World+BoT-SORT trajectory detector, and video-MLLM prompting. Across seven video generators, domain-adapted models (EnerVerse_FT, LTX_FT) consistently outperform commercial (Kling, Hailuo) and open-source (COSMOS, OpenSora, LTX) models, and EWMBench's rankings align more closely with human judgment than VBench. The work also shows that conventional perceptual metrics (e.g., VBench background consistency) are insensitive to viewpoint/layout instability critical for embodied tasks."},{"id":"arxiv:2308.16369","kind":"paper","n":246,"label":"Sarathi: Efficient LLM Inference by Piggybacking Decodes with Chunked Prefills","authors":"Amey Agrawal, Ashish Panwar, Jayashree Mohan, et al.","year":2023,"url":"https://arxiv.org/abs/2308.16369","region":"t24","hemi":"llm","lobe":"frontal","color":"#6784b0","indegree":3,"alsoIn":[],"x":0.8795,"y":0.6603,"text":"LLM inference has a compute-saturating prefill phase and a memory-bound decode phase that under-utilizes the GPU (decode cost-per-token can be ~200x prefill at batch size 1), and pipeline-parallel serving suffers bubbles from non-uniform micro-batch times. Sarathi introduces chunked-prefills (splitting a prompt into equal-sized compute chunks) and decode-maximal batching (one prefill chunk plus decodes piggybacked into remaining slots), fusing linear ops so decode tokens reuse already-fetched weights and become compute-bound. This yields up to 10x decode throughput and 1.33x end-to-end throughput for LLaMA-13B on A6000, and 4.25x decode / 1.25x end-to-end for LLaMA-33B on A100. On GPT-3 with pipeline parallelism, uniform hybrid batches cut median pipeline bubble time by 6.29x for a 1.91x end-to-end speedup."},{"id":"arxiv:2411.00769","kind":"paper","n":247,"label":"GameGen-X: Interactive Open-world Game Video Generation","authors":"Che et al.","year":2024,"url":"https://arxiv.org/abs/2411.00769","region":"t15","hemi":"wm","lobe":"occipital","color":"#4f9a62","indegree":2,"alsoIn":[],"x":0.287,"y":0.533,"text":"GameGen-X is the first diffusion transformer designed to both generate and interactively control open-world game videos, simulating game-engine features like novel characters, dynamic environments, actions, and events. The authors build OGameData, a 1-million-clip game video dataset (150+ next-gen games, GPT-4o structured captions) and train in two stages: pre-training a foundation MSDiT model for text-to-video generation and video continuation, then freezing it and training InstructNet for multi-modal interactive control via structured text instructions and keyboard inputs. On their OGameEval benchmark GameGen-X outperforms open-source video models on generation quality (FID 252.1, FVD 759.8) and substantially on control success rate (SR-C 63.0%, SR-E 56.8%). The work demonstrates generative models as a potential auxiliary tool to traditional game rendering."},{"id":"arxiv:2305.05383","kind":"paper","n":248,"label":"Code Execution with Pre-trained Language Models","authors":"Liu et al. (Chenxiao Liu, Shuai Lu, et al.)","year":2023,"url":"https://arxiv.org/abs/2305.05383","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":1,"alsoIn":[],"x":0.8484,"y":0.7255,"text":"Most pre-trained code models rely only on source code and syntactic structures (AST, data flow) and ignore the dynamic execution trace. This paper builds CodeExecutor, a 12-layer Transformer (initialized from UniXcoder) that is pre-trained on a new code-execution objective to predict a program's full execution trace \u2014 both the line execution order and intermediate variable states. To enable this, the authors create a large-scale Python dataset via mutation-based data augmentation over CodeNet submissions (plus SingleLine and Python Tutorial data) and train with curriculum learning from easy to hard. CodeExecutor substantially outperforms Codex on code execution and improves downstream zero-shot code-to-code search and text-to-code generation."},{"id":"arxiv:2509.19958","kind":"paper","n":249,"label":"Generalist Robot Manipulation beyond Action Labeled Data","authors":"","year":null,"url":"https://arxiv.org/abs/2509.19958","region":"t0","hemi":"wm","lobe":"frontal","color":"#9a4f4f","indegree":4,"alsoIn":[],"x":0.2201,"y":0.7687,"text":"Generalist VLA policies rely on scarce, expensive action-labeled robot demonstrations and degrade out-of-distribution. The paper proposes MotoVLA, a Mixture-of-Transformers VLA that pre-trains a '3D Dynamics Predictor' via flow matching on dense dynamic 3D point clouds extracted at the hand/gripper from unlabeled human and robot videos (RH20T, BridgeData V2, RT-1), then tunes it into an Action Predictor on a smaller action-labeled set (BridgeData V2) for action alignment. On the SIMPLER BridgeData V2 in-domain benchmark it reaches 68.2% average success, beating \u03c00 (B) by 11.4% and LAPA by 14.1%, and it enables 'out-of-action' generalization\u2014transferring skills from unlabeled human/robot demos to new robot actions on a real WidowX robot."},{"id":"arxiv:2605.16547","kind":"paper","n":250,"label":"World Model-Enabled Causal Digital Twins for Semantic Communications in Physical AI Systems","authors":"","year":null,"url":"https://arxiv.org/abs/2605.16547","region":"t39","hemi":"wm","lobe":"occipital","color":"#9a4f5b","indegree":3,"alsoIn":[],"x":0.2419,"y":0.573,"text":"The paper tackles goal-oriented semantic communication for closed-loop physical AI systems (e.g., UAVs) where transmitted semantics affect not just current inference but future control, state evolution, and long-horizon task return. It formulates the problem as long-term return-per-bit maximization under wireless bit-budget constraints, introduces a Causal Information Value (CIV) metric that uses counterfactual transmission interventions (Pearl's do-operator) to score each semantic token's marginal contribution to long-horizon return, and proposes a World-Model-Enabled Causal Digital Twin (WM-CDT) that learns RSSM latent dynamics at the edge server for imagined rollouts, actor-critic control, and CIV-guided token selection. On an AirSim-Sionna UAV navigation simulator with realistic wireless channel modeling, WM-CDT achieves up to 17.3%/55.4%/30.7%/33.7% higher return-per-kbit and 9.6%/26.3%/16.1%/18.0% higher navigation success rate than AC-RRL, MBPO, AC, and PPO baselines. CIV is also shown to correlate better with true long-horizon return gain than myopic VoI, saliency, and confidence scores."},{"id":"arxiv:2501.03575","kind":"paper","n":251,"label":"Cosmos World Foundation Model Platform for Physical AI","authors":"NVIDIA (Agarwal et al.)","year":2025,"url":"https://arxiv.org/abs/2501.03575","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":7,"alsoIn":[],"x":0.4393,"y":0.9496,"text":"NVIDIA presents the Cosmos World Foundation Model (WFM) Platform, an open-source suite for building world models for Physical AI under a pre-train-then-post-train paradigm. It contributes a scalable video curation pipeline (~100M clips distilled from 20M hours), a family of causal continuous/discrete video tokenizers (Cosmos-Tokenize1), and two families of pre-trained WFMs\u2014diffusion-based (7B/14B Text2World\u2192Video2World) and autoregressive (4B/12B\u21925B/13B)\u2014trained on 10,000 H100 GPUs over three months. Post-training examples adapt the WFMs to camera-controlled navigation, robotic manipulation, and multi-view autonomous driving, with a guardrail system for safe use. Cosmos tokenizers and WFMs achieve state-of-the-art reconstruction/3D-consistency and beat VideoLDM/CamCo baselines, though all models still struggle with strict physics adherence."},{"id":"arxiv:2411.10171","kind":"paper","n":252,"label":"Imagine-2-Drive: High-Fidelity World Modeling in CARLA for Autonomous Vehicles","authors":"Imagine-2-Drive authors","year":2024,"url":"https://arxiv.org/abs/2411.10171","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":9,"alsoIn":[],"x":0.3092,"y":0.6485,"text":"Imagine-2-Drive is a world-model-based reinforcement learning framework for autonomous driving in CARLA that addresses compounding one-step prediction errors and the limited expressiveness of deterministic/single-Gaussian RL policies. It couples DiffDreamer, a Stable Video Diffusion (SVD)-based world model (initialized from Vista) that predicts H future observations and rewards simultaneously, with DPA, a diffusion policy actor trained via PPO inside the world model's imagination to produce multi-modal waypoint trajectories. The two modules are trained alternately for stable co-evolution. On CARLA Town04 driving benchmarks it outperforms world-model and model-free baselines, improving Success Rate and Route Completion by ~20% and ~14.6% while halving infractions/km, and achieves far lower FID/FVD than prior world models."},{"id":"nianticlabs.com:05a236f27efe","kind":"paper","n":253,"label":"Building a Large Geospatial Model to Achieve Spatial Intelligence","authors":"Niantic Labs / Niantic Spatial","year":2024,"url":"https://nianticlabs.com/news/largegeospatialmodel","region":"t12","hemi":"llm","lobe":"frontal","color":"#5e9a4f","indegree":0,"alsoIn":[],"x":0.5732,"y":0.1731,"text":"Niantic presents a vision and position paper for a Large Geospatial Model (LGM): a foundation model that distills billions of geolocated images into a shared, metric-scale understanding of physical places. The core idea is to overcome the failure mode of independent local maps \u2014 which cannot position camera views from unseen angles \u2014 by sharing distilled knowledge across millions of scenes globally, so the model can infer unseen parts of a scene (e.g. the back of a church) from thousands of similar structures. It builds on Niantic's existing Visual Positioning System (VPS), which uses neural mapping methods (ACE, ACE Zero) and the two-view relative pose network MicKey as a proof of concept. The paper is conceptual, reporting deployment-scale statistics rather than benchmark results."},{"id":"arxiv:2407.06886","kind":"paper","n":254,"label":"Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AI","authors":"Yang Liu, Weixing Chen, Yongjie Bai, Xiaodan Liang, Guanbin Li, Wen Gao, Liang L","year":2024,"url":"https://arxiv.org/abs/2407.06886","region":"t7","hemi":"wm","lobe":"occipital","color":"#adb067","indegree":6,"alsoIn":[],"x":0.1958,"y":0.8054,"text":"This survey comprehensively reviews Embodied AI in the era of Multi-modal Large Models (MLMs) and World Models (WMs), framing it as the alignment of cyber space (simulation) with the physical world toward AGI. It organizes the field into embodied robots, simulators, and four research targets\u2014embodied perception, embodied interaction, embodied agents, and sim-to-real adaptation\u2014surveying state-of-the-art methods, paradigms, and datasets for each. As an original contribution, the authors propose the ARIO (All Robots In One) dataset standard and a unified large-scale dataset of ~3 million episodes from 258 series and 321,064 tasks. The paper introduces an ABC (AI brain, Body, Cross-modal sensors) framework for MLM/WM-based embodied agents and closes with challenges (data scarcity, long-horizon execution, causal reasoning, unified evaluation, security/privacy)."},{"id":"arxiv:2510.16732","kind":"paper","n":255,"label":"A Comprehensive Survey on World Models for Embodied AI","authors":"Li et al.","year":2025,"url":"https://arxiv.org/abs/2510.16732","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":11,"alsoIn":[],"x":0.2548,"y":0.892,"text":"This survey organizes the rapidly growing literature on world models for embodied AI under a unified framework. It formalizes world models as POMDP-based internal simulators trained with a reconstruction\u2013KL-regularization (ELBO) objective, and proposes a three-axis taxonomy: functionality (Decision-Coupled vs. General-Purpose), temporal modeling (Sequential Simulation/Inference vs. Global Difference Prediction), and spatial representation (Global Latent Vector, Token Feature Sequence, Spatial Latent Grid, Decomposed Rendering). It systematizes datasets and metrics across robotics, autonomous driving, and general video, and gives quantitative state-of-the-art comparisons, distilling open challenges around unified datasets, physics-aware (vs. pixel-fidelity) evaluation, real-time efficiency, and long-horizon consistency with error accumulation."},{"id":"arxiv:1810.04805","kind":"paper","n":256,"label":"BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding","authors":"Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova","year":2018,"url":"https://arxiv.org/abs/1810.04805","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":0,"alsoIn":[],"x":0.7946,"y":0.5178,"text":"BERT addresses the limitation that prior language-model pre-training (e.g., OpenAI GPT, ELMo) is unidirectional or only shallowly bidirectional, restricting representations for token- and sentence-level tasks. It pre-trains a deep bidirectional Transformer encoder using two unsupervised objectives \u2014 a masked language model (MLM) that predicts randomly masked tokens from both-side context, and next sentence prediction (NSP) \u2014 on BooksCorpus plus English Wikipedia (3.3B words). Fine-tuned with a single added output layer, BERT sets new state of the art on 11 NLP tasks, including pushing the GLUE score to 80.5% (+7.7 points), MNLI to 86.7%, SQuAD v1.1 Test F1 to 93.2, and SQuAD v2.0 Test F1 to 83.1. Ablations show the gains come primarily from deep bidirectionality enabled by the MLM objective."},{"id":"arxiv:2402.16823","kind":"paper","n":257,"label":"Language Agents as Optimizable Graphs","authors":"Zhuge et al.","year":2024,"url":"https://arxiv.org/abs/2402.16823","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":1,"alsoIn":[],"x":0.5513,"y":0.9366,"text":"GPTSwarm unifies disparate LLM-agent and prompt-engineering approaches by representing language agents as directed computational graphs, where nodes are operations (LLM queries, tool use) and edges define information flow; multiple agent graphs compose into a 'swarm' (composite graph). The paper introduces two automatic optimizers: edge optimization, which improves inter-agent orchestration/connectivity via a continuous relaxation optimized with REINFORCE, and node optimization, which iteratively refines per-node prompts using execution history. Experiments on MMLU, Mini Crosswords, HumanEval, and GAIA show the framework can build, integrate, and automatically improve agents\u2014e.g., filtering adversarial agents and recovering baseline accuracy, and surpassing prior methods like Tree-of-Thought. On GAIA it reaches 18.45% average vs 9.70% for GPT-4-Turbo."},{"id":"nature.com:e8ee4994d788","kind":"paper","n":258,"label":"Shared computational principles for language processing in humans and deep language models","authors":"Goldstein, A., Zada, Z., Buchnik, E., et al.","year":2022,"url":"https://www.nature.com/articles/s41593-022-01026-4","region":"t5","hemi":"llm","lobe":"parietal","color":"#b09e67","indegree":3,"alsoIn":[],"x":0.6375,"y":0.4835,"text":"The paper tests whether autoregressive deep language models (DLMs, specifically GPT-2) share computational principles with the human brain during natural language comprehension. Nine epilepsy patients listened to a 30-min podcast while ECoG recorded high-frequency broadband (70\u2013200 Hz) activity, and a separate 50-participant sliding-window behavioral study measured word-by-word prediction. Using linear encoding and deep nonlinear decoding models, the authors show the brain spontaneously predicts upcoming words before onset, calculates post-onset surprise tied to GPT-2's confidence, and represents words via context-dependent embeddings. They conclude autoregressive DLMs provide a biologically feasible framework for modeling the neural basis of language."},{"id":"arxiv:1911.08265","kind":"paper","n":259,"label":"Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model (MuZero)","authors":"Schrittwieser et al.","year":2020,"url":"https://arxiv.org/abs/1911.08265","region":"t22","hemi":"wm","lobe":"temporal","color":"#4f849a","indegree":3,"alsoIn":[],"x":0.273,"y":0.2267,"text":"Tree-based planning excels in games with a perfect simulator but fails in real-world domains where dynamics are unknown. MuZero combines Monte-Carlo tree search with a learned model that predicts only the three quantities relevant to planning \u2014 reward, value, and policy \u2014 via a representation, dynamics, and prediction function trained end-to-end, with no requirement to reconstruct observations or match true environment state. Without any knowledge of game rules or dynamics, MuZero achieved a new state of the art across 57 Atari games and matched AlphaZero's superhuman play in Go, chess, and shogi. A sample-efficient variant, MuZero Reanalyze, reached 731% median human-normalized Atari score in the 200M-frame regime."},{"id":"arxiv:2307.15054","kind":"paper","n":260,"label":"A Geometric Notion of Causal Probing","authors":"Cl\u00e9ment Guerner, Anej Svete, Tianyu Liu, Alexander Warstadt, Ryan Cotterell","year":2023,"url":"https://arxiv.org/abs/2307.15054","region":"t6","hemi":"llm","lobe":"temporal","color":"#b0a967","indegree":5,"alsoIn":[],"x":0.6914,"y":0.4636,"text":"The paper develops an intrinsic, information-theoretic framework for testing the linear subspace hypothesis \u2014 that a language model encodes a concept (e.g., verbal number) in a linear subspace of its representation space \u2014 using only distributions induced from the model rather than external probing classifiers. To avoid the spurious-correlation failure mode where a concept subspace and its orthogonal complement are correlated through a common cause (the context), the authors construct a counterfactual unigram distribution that forces the two subspaces to be statistically independent, and derive four intrinsic criteria (erasure, encapsulation, containment, stability) plus a causal graphical model that enables concept-controlled generation via a do-intervention. Empirically, R-LACE yields a 1-dimensional subspace that erases roughly half of total concept information while preserving non-concept information, and for GPT2-large a do-intervention on the concept recovers post-erasure accuracy back to above 0.9. Results are mixed for the French gender model, suggesting the causal intervention works precisely for at least one model/concept but not universally."},{"id":"arxiv:2006.11239","kind":"paper","n":261,"label":"Denoising Diffusion Probabilistic Models","authors":"Jonathan Ho, Ajay Jain, Pieter Abbeel","year":2020,"url":"https://arxiv.org/abs/2006.11239","region":"t32","hemi":"wm","lobe":"frontal","color":"#8b4f9a","indegree":5,"alsoIn":[],"x":0.3748,"y":0.0979,"text":"Diffusion probabilistic models had been defined but never shown to produce high-quality samples; this paper closes that gap. It introduces a simple \u03b5-prediction (noise-prediction) parameterization of the reverse process and establishes a novel equivalence between diffusion models and denoising score matching with annealed Langevin dynamics, yielding a simplified, reweighted variational training objective. The resulting models achieve state-of-the-art unconditional CIFAR10 sample quality (FID 3.17, Inception score 9.46) and sample quality on 256x256 LSUN comparable to ProgressiveGAN. The authors also show the reverse process is a progressive lossy decompression scheme that generalizes autoregressive decoding, with most lossless codelength spent on imperceptible detail."},{"id":"arxiv:2408.03943","kind":"paper","n":262,"label":"Building Machines that Learn and Think with People","authors":"Katherine M. Collins, Ilia Sucholutsky, Umang Bhatt, Kartik Chandra, Lionel Wong","year":2024,"url":"https://arxiv.org/abs/2408.03943","region":"t7","hemi":"llm","lobe":"occipital","color":"#adb067","indegree":3,"alsoIn":[],"x":0.6585,"y":0.5762,"text":"This Perspective argues that AI should move from being a tool for thought to a 'thought partner' that collaborates with humans, and proposes engineering such partners by having them explicitly build and reason over structured probabilistic models of the human, the world, and the task rather than relying solely on scaled foundation models trained to mimic behavior. It lays out modes of collaborative thought (planning, learning, deliberation, sensemaking, creation), three desiderata ('you understand me,' 'I understand you,' 'we understand the world'), and a 'Bayesian Thought Partner Toolkit' of nine computational cognitive science motifs. The authors ground the proposal in case studies (WatChat for programming, CLIPS for embodied assistance, inverse inverse planning for storytelling, probabilistic generative models for medicine) and discuss infrastructure, evaluation, and risks. The contribution is conceptual/programmatic \u2014 a research agenda and framework, not empirical benchmark results."},{"id":"arxiv:2403.16971","kind":"paper","n":263,"label":"AIOS: LLM Agent Operating System","authors":"Mei et al.","year":2024,"url":"https://arxiv.org/abs/2403.16971","region":"t35","hemi":"llm","lobe":"occipital","color":"#b0679e","indegree":1,"alsoIn":[],"x":0.6283,"y":0.9432,"text":"LLM-based agents lack proper resource management, letting them monopolize LLM/tool resources and forcing inefficient trial-and-error GPU loading under concurrency. AIOS proposes an agent operating system that isolates resources and LLM-specific services from agent applications into an 'AIOS kernel' providing scheduling, context, memory, storage, tool, and access management, plus an SDK to access these via system calls. Agent queries are decomposed into thread-bound syscalls dispatched by a scheduler (FIFO/Round Robin) with a context-interrupt mechanism for preemptive LLM inference. Experiments across five agent frameworks show AIOS preserves or improves benchmark accuracy and achieves up to 2.1\u00d7 faster execution/throughput while scaling near-linearly to 2000 concurrent agents."},{"id":"arxiv:2504.10612","kind":"paper","n":264,"label":"Energy Matching: Unifying Flow Matching and Energy-Based Models for Generative Modeling","authors":"Michal Balcerak, Tamaz Amiranashvili, Suprosanna Shit, Antonio Terpin, Sebastian","year":2025,"url":"https://arxiv.org/abs/2504.10612","region":"t3","hemi":"wm","lobe":"occipital","color":"#9a714f","indegree":2,"alsoIn":[],"x":0.3144,"y":0.1205,"text":"The paper introduces Energy Matching, a generative framework that unifies flow matching and energy-based models (EBMs) by learning a single time-independent scalar potential V_\u03b8(x). Far from the data manifold the gradient field drives irrotational, optimal-transport-like flow from noise to data (trained simulation-free), while near the manifold an entropic term turns the dynamics into Langevin sampling of a Boltzmann equilibrium \u03c1 \u221d exp(\u2212V_\u03b8/\u03b5_max) trained by contrastive divergence. The method substantially beats prior EBMs on CIFAR-10 (FID 3.34) and ImageNet 32\u00d732 (FID 6.64) without auxiliary generators or time conditioning, and its explicit likelihood enables inverse-problem solving, repulsion-based diverse protein design, and more accurate local-intrinsic-dimension estimation."},{"id":"arxiv:2105.11203","kind":"paper","n":265,"label":"How particular is the physics of the free energy principle?","authors":"Miguel Aguilera, Beren Millidge, Alexander Tschantz, Christopher L. Buckley","year":2022,"url":"https://arxiv.org/abs/2105.11203","region":"t25","hemi":"llm","lobe":"parietal","color":"#6779b0","indegree":6,"alsoIn":[],"x":0.5987,"y":0.8232,"text":"The paper presents a mathematical and conceptual critique of the free energy principle (FEP) by testing the assumptions needed to derive it in the simplest tractable non-equilibrium case: weakly-coupled linear stochastic (Ornstein-Uhlenbeck/Langevin) systems. Using power-series (Neumann) expansions of the steady-state covariance, Hessian and solenoidal-flow matrices in the coupling strength, the authors show that the Markov blanket condition and the required absence of inter-block solenoidal couplings hold only for a very narrow, highly symmetric parameter space that excludes the perception-action asymmetries expected in living systems. They further argue that a central step \u2014 equating the rate of change of average states with the average (marginal) flow conditioned on blanket states \u2014 is invalid once stochastic fluctuations and the system's history of interactions are present, even for linear systems. The conclusion is that the FEP, as currently formulated, is not straightforwardly applicable even to these simple systems, undermining claims of its generality."},{"id":"arxiv:2602.23152","kind":"paper","n":266,"label":"The Trinity of Consistency as a Defining Principle for General World Models","authors":"","year":null,"url":"https://arxiv.org/abs/2602.23152","region":"t6","hemi":"wm","lobe":"temporal","color":"#b0a967","indegree":0,"alsoIn":[],"x":0.2911,"y":0.7314,"text":"This survey proposes the 'Trinity of Consistency' \u2014 Modal (semantic interface), Spatial (geometric basis), and Temporal (causal engine) consistency \u2014 as a defining theoretical framework for what constitutes a General World Model, and uses it to review the evolution of generative/multimodal models from loosely coupled specialized modules toward unified architectures. To operationalize the framework, the authors introduce CoW-Bench, a 1,485-sample benchmark with 18 sub-tasks (6 task families covering the three single consistencies and their three pairwise intersections) evaluated as a constraint-satisfaction problem via 16 reusable atomic checks and human checklists. Results show closed-source image generators dominate (GPT-image-1.5 scores 85.62 average, Nano Banana Pro 82.57) while cross-consistency 'fusion' tasks like navigation/maze (TS-Maze-2D) remain unsolved even for top models, exposing a 'constraint-backoff' failure where models produce plausible pixels while silently violating logical commitments."},{"id":"arxiv:2506.08967","kind":"paper","n":267,"label":"Step-Audio-AQAA: a Fully End-to-End Expressive Large Audio Language Model","authors":"Ailin Huang et al.","year":2025,"url":"https://arxiv.org/abs/2506.08967","region":"t11","hemi":"llm","lobe":"occipital","color":"#81b067","indegree":7,"alsoIn":[],"x":0.8798,"y":0.266,"text":"Step-Audio-AQAA is a fully end-to-end Large Audio-Language Model for Audio Query-Audio Answer (AQAA) tasks that takes raw audio in and emits natural speech directly, without separate ASR or TTS stages. It combines a dual-codebook audio tokenizer (linguistic + semantic), a 130B-parameter multimodal backbone LLM (Step-Omni), and a CosyVoice-style flow-matching neural vocoder, post-trained via two-stage SFT, audio-token-masked DPO, and weight merging, with interleaved text+audio token output. On the StepEval-Audio-360 human-MOS benchmark it outperforms Kimi-Audio and Qwen-Omni in speech emotion control, creativity, role-playing, language ability, gaming, logical reasoning, and voice understanding, while lagging in singing and voice-instruction-following. Ablations show that interleaving text and audio tokens at a 10:15 ratio and using marker-preserving concatenation gives the best chat/relevance/factuality scores."},{"id":"the-decoder.com:97885f462d1f","kind":"paper","n":268,"label":"LLMs could serve as world models for training AI agents, study finds","authors":"The Decoder","year":2026,"url":"https://the-decoder.com/llms-could-serve-as-world-models-for-training-ai-agents-study-finds/","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":18,"alsoIn":[],"x":0.7034,"y":0.792,"text":"The paper investigates whether large language models can serve as world models \u2014 internal simulators that predict the next environment state given an action \u2014 to generate synthetic experience for training autonomous AI agents, addressing the bottleneck of limited real-environment interaction. The authors reframe language modeling as next-state (rather than next-token) prediction and evaluate across five text-based environments on transition-prediction accuracy, scaling behavior, and usefulness for agent training. Pre-trained models show partial ability (Claude-sonnet-4.5 reached 77% next-state accuracy on ALFWorld with three examples), but fine-tuning on real interaction data unlocks high-fidelity simulation (>99% on ALFWorld). The work provides empirical support for experience-based agent training while noting it does not solve continuous learning without forgetting."},{"id":"arxiv:2410.13056","kind":"paper","n":269,"label":"Channel-Wise Mixed-Precision Quantization for Large Language Models","authors":"","year":null,"url":"https://arxiv.org/abs/2410.13056","region":"t37","hemi":"llm","lobe":"parietal","color":"#b06788","indegree":0,"alsoIn":[],"x":0.8695,"y":0.4066,"text":"LLMs are hard to deploy on edge devices because of their memory footprint, and existing post-training weight-only quantization is restricted to integer bit-widths, wasting available storage when devices could support fractional precision (e.g., 2.2 bits). The paper proposes CMPQ, a channel-wise mixed-precision PTQ method that assigns higher/lower precision to weight channels based on per-channel activation L2-norms, uses non-uniform (K-means) quantization, and protects two kinds of outliers (activation-based and quantization-aware) in FP16. Across nine OPT/LLaMA2/LLaMA3 models, CMPQ matches or beats integer-bit baselines (GPTQ, AWQ, QuIP, QuIP#, LLM-MQ, SliM-LLM) and uniquely supports any average bit-width, e.g. improving LLaMA2-7B C4 perplexity ~30% (15.97\u219211.11) when moving from 2 to 2.2 bits. It does this using only forward passes (no backprop), so it needs roughly half the quantization-time memory of gradient-based SqueezeLLM while remaining within ~1% of its accuracy."},{"id":"arxiv:2504.10903","kind":"paper","n":270,"label":"Efficient Reasoning Models: A Survey","authors":"Sicheng Feng et al.","year":2025,"url":"https://arxiv.org/abs/2504.10903","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":4,"alsoIn":[],"x":0.8335,"y":0.8161,"text":"This survey organizes research on efficient reasoning for Large Reasoning Models (LRMs) that generate long Chain-of-Thoughts and thereby incur heavy compute overhead. It proposes a goal-driven taxonomy of three directions: 'shorter' (compressing lengthy CoTs via RL length penalties, variable-length-CoT SFT, prompting/routing, and latent reasoning), 'smaller' (building compact reasoning models via distillation, pruning/quantization, and RL), and 'faster' (efficient decoding via efficient sampling, efficient self-consistency, and problem decomposition). It documents the overhead concretely (e.g., DeepSeek-R1 uses 619 tokens to answer '1 plus 2'; 15513 tokens for the 1.5B model on AIME 24) and compiles comparative tables of methods, metrics, datasets, and benchmarks."},{"id":"arxiv:2107.08241","kind":"paper","n":271,"label":"High-Accuracy Model-Based Reinforcement Learning, a Survey","authors":"Aske Plaat, Walter Kosters, Mike Preuss","year":2021,"url":"https://arxiv.org/abs/2107.08241","region":"t26","hemi":"llm","lobe":"temporal","color":"#4f579a","indegree":9,"alsoIn":[],"x":0.5897,"y":0.8431,"text":"This survey reviews deep model-based reinforcement learning methods that aim to achieve high model accuracy at low sample complexity for high-dimensional sequential decision problems. It presents a taxonomy along three axes\u2014application type (continuous/discrete action spaces), learning method (uncertainty modeling, ensembles, latent models, CNN/LSTM), and planning method (trajectory rollouts, MPC, end-to-end learning and planning)\u2014and traces the influence of key methods from PILCO through MuZero. The survey concludes that model-based methods now approach model-free accuracy with substantially fewer environment samples, but notes challenges in reproducibility, hyperparameter sensitivity, and applicability to new domains, proposing a six-item research agenda."},{"id":"arxiv:2507.06952","kind":"paper","n":272,"label":"What Has a Foundation Model Found? Using Inductive Bias to Probe for World Models","authors":"Vafa, Chang, Rambachan & Mullainathan","year":2025,"url":"https://arxiv.org/abs/2507.06952","region":"t6","hemi":"wm","lobe":"temporal","color":"#b0a967","indegree":0,"alsoIn":[],"x":0.273,"y":0.5606,"text":"The paper develops the 'inductive bias probe,' a technique that evaluates whether a foundation model has internalized a postulated world model by repeatedly adapting it to small synthetic datasets consistent with that world model and measuring whether its extrapolations align with the world model's state structure. Applied across orbital mechanics, lattice navigation, and Othello, the method shows that models can excel at next-token prediction yet have weak inductive bias toward the true underlying world model. A transformer trained on orbital trajectories predicts positions almost perfectly (R\u00b2>0.9999) but, when fine-tuned to predict forces, recovers nonsensical, sample-dependent gravitational laws rather than Newton's law. Further analysis shows models behave as if they build task-specific heuristics (e.g. grouping states by their legal next-token sets) instead of coherent world models."},{"id":"arxiv:2404.07214","kind":"paper","n":273,"label":"Exploring the Frontier of Vision-Language Models: A Survey of Current Methodologies and Future Directions","authors":"","year":null,"url":"https://arxiv.org/abs/2404.07214","region":"t30","hemi":"llm","lobe":"temporal","color":"#8c67b0","indegree":7,"alsoIn":[],"x":0.7589,"y":0.205,"text":"This paper is a comprehensive survey of approximately 70 Vision-Language Models (VLMs), organized into three categories based on input-output modality capabilities: Vision-Language Understanding models, Text Generation with Multimodal Input models, and Multimodal Output with Multimodal Input models. For each model, the authors analyze architecture, training data, strengths, and limitations, and provide comparative performance across ten benchmark datasets (Table 1), the MME benchmark (Table 2), and video QA datasets (Table 3). The survey also outlines future directions including modular architectures, additional modalities, fine-grained evaluation, causality, continual learning, training efficiency, multilingual grounding, and domain-specific VLMs."},{"id":"arxiv:2510.07092","kind":"paper","n":274,"label":"Generative World Modelling for Humanoids: 1X World Model Challenge Technical Report","authors":"Team Revontuli (1X World Model Challenge)","year":2025,"url":"https://arxiv.org/abs/2510.07092","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":1,"alsoIn":[],"x":0.3231,"y":0.843,"text":"This technical report describes Team Revontuli's winning entries in the 1X World Model Challenge, a benchmark of real-world humanoid interaction with two tracks: sampling (forecasting future RGB frames) and compression (predicting future discrete latent token grids). For sampling, they adapt the Wan-2.2 TI2V-5B flow-matching video foundation model to video-and-robot-state-conditioned future-frame prediction via AdaLN-Zero state conditioning and LoRA fine-tuning; for compression they train a Spatio-Temporal Transformer from scratch on Cosmos-tokenised video. The models reach 23.0 dB PSNR (sampling) and a Top-500 cross-entropy of 6.6386 (compression), placing 1st in both tracks. Ensemble averaging (selective blurring of high-uncertainty regions) for sampling and greedy autoregressive decoding for compression were the most effective inference strategies."},{"id":"arxiv:2606.00133","kind":"paper","n":275,"label":"World Models: A Comprehensive Survey of Architectures, Methodologies, Reasoning Paradigms, and Applications","authors":"Zidan, Pan, Jiang et al. (26 authors)","year":2026,"url":"https://arxiv.org/abs/2606.00133","region":"t17","hemi":"llm","lobe":"parietal","color":"#4f9a79","indegree":29,"alsoIn":[],"x":0.6794,"y":0.6285,"text":"This survey addresses the lack of a unified framework for the rapidly fragmenting world-model field by introducing a comprehensive multi-axis taxonomy organized along four dimensions: architecture (representation, dynamics, modality, learning paradigm, downstream use), methodological family (state-space/recurrent, transformer, diffusion, physics-informed, language-augmented), reasoning strategy (imagination-based planning, latent policy learning, counterfactual reasoning, planning under uncertainty), and application domain. It traces the field from cognitive-science foundations and Ha & Schmidhuber's neural world models through milestone systems (PlaNet, Dreamer, MuZero, Sora, Cosmos, Genie, V-JEPA 2) and highlights the emerging convergence of chain-of-thought reasoning with world-model imagination. The authors review evaluation protocols/benchmarks and identify persistent challenges\u2014compounding prediction errors, sim-to-real transfer, and fragmented evaluation\u2014while extending coverage to underexplored domains such as medical imaging, educational measurement, and finance. They conclude with future directions toward unified multimodal world models, foundation-scale interactive simulators, and safe deployment, plus a proposed distinction between predictive world models and generative world simulators."},{"id":"arxiv:2102.08363","kind":"paper","n":276,"label":"COMBO: Conservative Offline Model-Based Policy Optimization","authors":"Yu et al.","year":2021,"url":"https://arxiv.org/abs/2102.08363","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":5,"alsoIn":[],"x":0.1759,"y":0.2937,"text":"Offline model-based RL methods like MOPO and MOReL rely on explicit uncertainty quantification for conservatism, which is unreliable with deep neural networks. COMBO instead learns a conservative value function by penalizing Q-values on out-of-support state-action pairs generated via model rollouts, while training on both offline data and synthetic model data, eliminating the need for uncertainty estimation. Theoretically, COMBO provides a lower bound on the true Q-function and a safe policy improvement guarantee. Empirically, COMBO achieves the best score in 9 out of 12 D4RL MuJoCo domains, outperforms prior methods on generalization tasks, and scales to image-based robotic manipulation."},{"id":"arxiv:2111.09259","kind":"paper","n":277,"label":"Acquisition of Chess Knowledge in AlphaZero","authors":"Thomas McGrath, Andrei Kapishnikov, Nenad Toma\u0161ev, Adam Pearce, Demis Hassabis, ","year":2021,"url":"https://arxiv.org/abs/2111.09259","region":"t5","hemi":"llm","lobe":"parietal","color":"#b09e67","indegree":2,"alsoIn":[],"x":0.6465,"y":0.7564,"text":"This interpretability study probes the AlphaZero chess network (a 20-block ResNet, 8\u00d78\u00d7256 = 16,384 activations per layer, trained over 1,000,000 self-play steps) to test whether a superhuman, self-play agent that never saw human games nonetheless acquires human chess concepts. Using sparse linear/logistic probes (concept activation vectors) for 93 Stockfish 8 evaluation sub-functions plus 116 custom concepts, the authors build 'what-when-where' plots showing which concepts are linearly decodable, when in training they emerge, and where in the network they are computed. They find many human concepts are accurately regressed (e.g. Stockfish total score r\u00b2 > 0.75 by block 10 at \u226564k steps), that value-function reliance shifts from material early to king-safety/mobility later, and that AlphaZero's opening evolution and skill acquisition (tactics before positional play) follow a rapid early 'critical period' around 25k\u201360k steps. Unsupervised NMF and activation-input covariance analyses corroborate that move/threat computations develop progressively across layers."},{"id":"arxiv:2209.15571","kind":"paper","n":278,"label":"Building Normalizing Flows with Stochastic Interpolants","authors":"Michael S. Albergo, Eric Vanden-Eijnden","year":2022,"url":"https://arxiv.org/abs/2209.15571","region":"t3","hemi":"wm","lobe":"occipital","color":"#9a714f","indegree":2,"alsoIn":[],"x":0.3202,"y":0.1103,"text":"The paper proposes a generative model that builds continuous-time normalizing flows between arbitrary base and target densities using a 'stochastic interpolant' \u2014 a time-dependent process constructed by independently sampling from both densities and passing them through an interpolant function. The velocity field of the associated probability flow ODE is learned as the minimizer of a simple quadratic objective, avoiding the costly backpropagation through ODE solvers required by conventional maximum-likelihood flow training. The method (InterFlow) matches or surpasses continuous flow baselines on tabular density estimation at a fraction of the training cost, achieves competitive NLL and FID on CIFAR-10 and ImageNet 32\u00d732 image generation, and scales ab-initio ODE flows to 128\u00d7128 image resolution \u2014 previously unreachable under MLE training."},{"id":"arxiv:2402.03161","kind":"paper","n":279,"label":"Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization","authors":"Yang Jin, Zhicheng Sun, Kun Xu, et al. (PKU, Kuaishou)","year":2024,"url":"https://arxiv.org/abs/2402.03161","region":"t29","hemi":"llm","lobe":"parietal","color":"#6a4f9a","indegree":2,"alsoIn":[],"x":0.7416,"y":0.2238,"text":"The paper tackles efficient video-language pre-training for multimodal LLMs, where encoding spatiotemporal dynamics normally requires prohibitively many tokens. Video-LaVIT decomposes each video into keyframes (tokenized by an off-the-shelf LaVIT image tokenizer, ~90 tokens) and temporal motion vectors extracted via MPEG-4 compression, which a VQ-VAE spatiotemporal motion tokenizer discretizes into just 135 tokens per 24-frame clip, enabling unified autoregressive next-token pre-training over video, image, and text with a Llama 2 7B backbone. A sequential video detokenizer (keyframe U-Net + 3D U-Net with enhanced motion conditioning) plus a DDIM-inversion noise constraint recovers pixel-space video and supports temporally consistent long-video generation. Across 13 benchmarks it achieves competitive image/video understanding and zero-shot text-to-image/video generation, e.g. SOTA on several image QA and video QA sets and leading FVD on MSR-VTT."},{"id":"arxiv:2605.06192","kind":"paper","n":280,"label":"EA-WM: Event-Aware Generative World Model with Structured Kinematic-to-Visual Action Fields","authors":"Zhaoyang Yang, Yurun Jin, Lizhe Qi, Cong Huang, Kai Chen","year":2026,"url":"https://arxiv.org/abs/2605.06192","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":9,"alsoIn":[],"x":0.305,"y":0.8579,"text":"Recent world-action models treat video generation as an auxiliary representation for policy learning and underexplore the inverse problem of using action signals to guide accurate video synthesis, often failing to preserve robot geometry and robot-object interaction dynamics. The paper presents EA-WM, an event-aware generative world model built on the Wan2.2-TI2V video diffusion backbone that projects low-dimensional robot actions and kinematic states into the target camera view as Structured Kinematic-to-Visual Action Fields (KVAFs) and fuses them with the video stream via event-aware bidirectional fusion blocks supervised by Event-Difference Latent Supervision (EDLS). Evaluated on the WorldArena benchmark, EA-WM reaches a P3CScore of 76.60, beating the strongest baseline (CogVideoX, 71.08) by 5.52 points and winning 5 of 6 selected metrics. Ablations confirm KVAFs and event-aware fusion contribute 5.63 and 1.80 P3CScore points respectively, and a KVAF-conditioned variant reaches 78.13 P3CScore."},{"id":"arxiv:2311.15127","kind":"paper","n":281,"label":"Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets","authors":"Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilia","year":2023,"url":"https://arxiv.org/abs/2311.15127","region":"t32","hemi":"wm","lobe":"frontal","color":"#8b4f9a","indegree":6,"alsoIn":[],"x":0.4492,"y":0.3292,"text":"Stable Video Diffusion (SVD) is a latent video diffusion model for high-resolution text-to-video and image-to-video generation that systematizes the role of data curation in training video LDMs. The authors identify three training stages \u2014 text-to-image pretraining, large-scale video pretraining at low resolution, and high-quality video finetuning \u2014 and build a curation pipeline (cut detection, optical-flow/OCR/CLIP-aesthetics filtering, synthetic captioning) that turns a raw 580M-clip collection (LVD) into a filtered 152M-clip dataset (LVD-F). Curated pretraining yields models preferred by human evaluators that beat WebVid-10M/InternVid-trained baselines and outperform closed-source Gen-2 and PikaLabs on image-to-video. The base model serves as a strong motion and multi-view 3D prior: finetuned into SVD-MV it beats Zero123XL and SyncDreamer on novel-view synthesis at a fraction of their compute."},{"id":"arxiv:2506.09849","kind":"paper","n":282,"label":"IntPhys 2: Benchmarking Intuitive Physics Understanding In Complex Synthetic Environments","authors":"Bordes, Garrido, Kao, Williams, Rabbat, Dupoux","year":2025,"url":"https://arxiv.org/abs/2506.09849","region":"t38","hemi":"llm","lobe":"temporal","color":"#9a4f66","indegree":3,"alsoIn":[],"x":0.6854,"y":0.3186,"text":"IntPhys 2 is a video benchmark that evaluates intuitive physics understanding in deep learning models using the violation-of-expectation paradigm, covering four core principles for macroscopic objects: Permanence, Immutability, Spatio-Temporal Continuity, and Solidity. Built in Unreal Engine with photorealistic scenes, dynamic shadows/lighting, occlusions, and both fixed and moving cameras, it contains 1,416 videos arranged in matched quadruplets (two possible, two impossible) across Debug, Main (Easy/Medium/Hard), and Held-Out splits. Evaluating state-of-the-art MLLMs (GPT-4o, Qwen-VL 2.5, Gemini 1.5 Pro, Gemini 2.5 Flash) and predictive models (VideoMAEv2, Cosmos-Predict 4B, V-JEPA, V-JEPA 2), the authors find nearly all models perform near chance (50%), while humans reach ~96% overall. The result exposes a large gap between current models and human-like intuitive physics, especially under occlusion and short-term memory demands in complex scenes."},{"id":"arxiv:2410.24164","kind":"paper","n":283,"label":"\u03c00: A Vision-Language-Action Flow Model for General Robot Control","authors":"Black, Brown, Driess, Esmail, Equi et al. (Physical Intelligence)","year":2024,"url":"https://arxiv.org/abs/2410.24164","region":"t0","hemi":"wm","lobe":"frontal","color":"#9a4f4f","indegree":0,"alsoIn":[],"x":0.148,"y":0.7693,"text":"\u03c00 is a generalist robot foundation model (vision-language-action policy) that adds a flow-matching action expert on top of the pre-trained PaliGemma VLM to generate continuous, high-frequency (up to 50 Hz) action chunks for dexterous manipulation. It is pre-trained on a large cross-embodiment mixture \u2014 over 10,000 hours of in-house data from 7 robot configurations and 68 tasks plus the OXE open-source datasets \u2014 then post-trained/fine-tuned on curated high-quality data for specific downstream tasks. Across out-of-box, language-following, and fine-tuning evaluations it substantially outperforms baselines such as OpenVLA, Octo, ACT, and Diffusion Policy, with near-perfect success on shirt folding and easy table bussing. It solves very long-horizon dexterous tasks (5\u201320 min) including laundry folding, table bussing, box assembly, and packing eggs."},{"id":"arxiv:2502.13092","kind":"paper","n":284,"label":"Text2World: Benchmarking Large Language Models for Symbolic World Model Generation","authors":"Hu, Chen, Wang et al.","year":2025,"url":"https://arxiv.org/abs/2502.13092","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":2,"alsoIn":[],"x":0.6809,"y":0.9163,"text":"The paper introduces TEXT2WORLD, a benchmark for evaluating whether LLMs can generate symbolic world models (PDDL domain definitions) from natural-language descriptions, addressing prior work's limited domain scope, evaluation randomness, and indirect metrics. It curates 103 diverse PDDL domains (from 1,801 raw files) with human-annotated abstract descriptions and proposes multi-criteria, execution-based metrics (executability, structural similarity, component-wise F1). Benchmarking 16 LLMs from 9 families shows reasoning models trained with large-scale RL (e.g., DeepSeek-R1) lead but even the best still struggles, and the authors explore test-time scaling, in-context learning, fine-tuning, and agent training as enhancement strategies."},{"id":"arxiv:2505.24631","kind":"paper","n":285,"label":"Cascades on Constrained Multiplex Networks","authors":"","year":null,"url":"https://arxiv.org/abs/2505.24631","region":"t36","hemi":"llm","lobe":"frontal","color":"#b06793","indegree":3,"alsoIn":[],"x":0.6151,"y":0.7534,"text":"The paper studies a Watts-type complex-contagion cascade model (with activation threshold fixed at \u03d5=1, using the 'or'-rule) on directed multiplex configuration-model networks with arbitrary joint degree distributions, deriving message-passing equations for expected cascade size, single-seed cascade probability (via a multi-type branching process over 'vulnerable links'), and two equivalent cascade conditions. It then introduces 'constrained multiplex networks', a tractable subclass in which node-activity patterns are set explicitly through a constraint matrix C, giving the clean cascade condition |\u03bbC|\u00b7Pin(1) > 1. Through analysis and simulation the authors show that induced node-activity patterns reshape the model's phase transitions: a normally continuous cascade-onset transition can become explosive (first-order), new nested cascade regions and a 'central cusp' transition appear, and weakly connected networks produce multiple distinct single-seed cascade sizes. A key negative result is that the standard \u03c10\u21920 macroscopic method can fail to predict single-seed cascade sizes in networks with multiple giant strongly connected components."},{"id":"arxiv:2106.00737","kind":"paper","n":286,"label":"Implicit Representations of Meaning in Neural Language Models","authors":"Li, Nye & Andreas","year":2021,"url":"https://arxiv.org/abs/2106.00737","region":"t6","hemi":"llm","lobe":"temporal","color":"#b0a967","indegree":4,"alsoIn":[],"x":0.7258,"y":0.6455,"text":"The paper investigates whether neural language models trained on text alone represent the situations they describe, not just surface co-occurrence statistics. Using frozen BART and T5 encoders fine-tuned on the Alchemy and TextWorld domains, the authors train linear/bilinear probes to recover 'information states' \u2014 the truth values of logical propositions about entities \u2014 from contextual token embeddings. They find entity state is linearly decodable, primarily attributable to open-domain pretraining rather than in-domain fine-tuning, localized to entity mentions, and causally manipulable: editing encoder representations changes generation in predictable ways. The conclusion is that LM prediction is supported in part by implicit, dynamic simulation of entity state learned from text only."},{"id":"arxiv:2310.16828","kind":"paper","n":287,"label":"TD-MPC2: Scalable, Robust World Models for Continuous Control","authors":"Hansen, Su & Wang","year":2024,"url":"https://arxiv.org/abs/2310.16828","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":0,"alsoIn":[],"x":0.1848,"y":0.2928,"text":"TD-MPC2 is a model-based RL algorithm that performs local trajectory optimization (MPPI planning) in the latent space of a learned implicit, decoder-free world model, improving on TD-MPC with more robust design choices (SimNorm latent normalization, discrete reward/value regression in log space, maximum-entropy policy prior, Q-function ensemble) so a single set of hyperparameters works across all tasks. Evaluated on 104 continuous control tasks spanning DMControl, Meta-World, ManiSkill2, and MyoSuite, it consistently outperforms SAC, DreamerV3, and TD-MPC in data-efficiency and asymptotic performance. The authors show capabilities scale with model size, training a single 317M-parameter agent on 80 tasks across multiple embodiments and action spaces, and release 300+ checkpoints, datasets, and code."},{"id":"arxiv:1801.04819","kind":"paper","n":288,"label":"Robots as Powerful Allies for the Study of Embodied Cognition from the Bottom Up","authors":"","year":null,"url":"https://arxiv.org/abs/1801.04819","region":"t7","hemi":"llm","lobe":"occipital","color":"#adb067","indegree":0,"alsoIn":[],"x":0.657,"y":0.5028,"text":"A theoretical/review chapter arguing that robots are the tools of choice for studying embodied cognition from the bottom up, combining a robotic approach with information theory and a developmental perspective. It surveys a progression from control-free 'low-level' behaviors (passive dynamic walkers, self-stabilizing mechanics) through reflex-based sensorimotor loops (Braitenberg vehicles, Brooks' subsumption architecture) to 'minimal cognition' and sensorimotor contingencies (SMCs), then to humanoid and developmental robotics. Its central concrete demonstration uses the quadruped robot 'Puppy' with transfer entropy to quantify how body, gait, and environment shape the sensorimotor space, showing that action context (gait) improves perceptual categorization such as terrain classification."},{"id":"arxiv:2605.16395","kind":"paper","n":289,"label":"OrbiSim: World Models as Differentiable Physics Engines for Embodied Intelligence","authors":"NeoWorld Project authors","year":2026,"url":"https://arxiv.org/abs/2605.16395","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":1,"alsoIn":[],"x":0.2108,"y":0.7245,"text":"OrbiSim reframes a learned world model as a fully differentiable physics engine for embodied manipulation, decoupling the neural architecture into OrbiSim-Dynamics (an object-centric recurrent state-space model with a Transformer coupling module that predicts explicit next physical states) and OrbiSim-Vision (a state-guided latent diffusion renderer). End-to-end differentiability spans state transitions through pixel generation, enabling differentiable contact modeling, Real-to-Sim system identification (via StaInf/PhyInf inference modules), and analytical policy gradients under sparse rewards. On a robosuite Push benchmark it cuts FVD to 533.9 vs 1305.8 (AdaWorld) and 1750.1 (Vid2World) and raises PSNR100 to 19.98 vs 12.83, and as an RL engine reaches a 42.71% task success rate vs 25.00% for DreamerV3 and \u22642.08% for model-free baselines. It also generalizes to articulated (AdaManip) and deformable (Physion Drape) objects under a single asset-conditioned interface."},{"id":"nature.com:a4d8ac562091","kind":"paper","n":290,"label":"Brains and algorithms partially converge in natural language processing","authors":"Charlotte Caucheteux, Jean-Remi King","year":2022,"url":"https://www.nature.com/articles/s42003-022-03036-1","region":"t28","hemi":"llm","lobe":"frontal","color":"#7667b0","indegree":8,"alsoIn":[],"x":0.648,"y":0.4732,"text":"The paper investigates what computational principles make deep language models produce brain-like representations by comparing a large set of trained-from-scratch transformers against human brain responses. Using an open dataset of 102 subjects who read 400 isolated Dutch sentences while recorded with both fMRI and source-localized MEG, the authors fit linear mappings (brain scores) from model activations to brain signals across 32,400 embeddings spanning architecture, training amount, and task. They find that the model-brain similarity depends primarily on a model's ability to predict words from context, and that this mapping recovers a spatiotemporal hierarchy of visual, lexical, and compositional representations. They conclude that modern language algorithms partially converge toward brain-like solutions, with middle layers and word-prediction performance being the dominant drivers."},{"id":"arxiv:2507.12440","kind":"paper","n":291,"label":"EgoVLA: Learning Vision-Language-Action Models from Egocentric Human Videos","authors":"Yang, Chalvatzaki, et al.","year":2025,"url":"https://arxiv.org/abs/2507.12440","region":"t0","hemi":"llm","lobe":"frontal","color":"#9a4f4f","indegree":6,"alsoIn":[],"x":0.5822,"y":0.213,"text":"EgoVLA trains a Vision-Language-Action model on large-scale egocentric human manipulation videos to predict human wrist poses and MANO hand parameters, then converts these to robot actions via inverse kinematics and retargeting, with a small amount of robot demonstration fine-tuning to obtain a bimanual humanoid policy. The authors build a ~500,000 image-action-pair egocentric dataset from four sources (HOI4D, HOT3D, HoloAssist, TACO) and introduce the Ego Humanoid Manipulation Benchmark in NVIDIA Isaac Lab with 12 bimanual tasks (Unitree H1 + Inspire hands) and 100 demos per task. Human-video pretraining yields large gains over a no-pretrain baseline and a per-task ACT specialist, especially on long-horizon and unseen-background tasks. The paper shows human pretraining boosts both in-domain success and out-of-domain generalization, though robot-data fine-tuning remains necessary (zero-shot deployment gives 0% success)."},{"id":"openaipublic.blob.core.windows.net:570a1f9c6e10","kind":"paper","n":292,"label":"Language Models Can Explain Neurons in Language Models","authors":"Bills et al. (OpenAI: Steven Bills, Nick Cammarata, Dan Mossing, Henk Tillman, e","year":2023,"url":"https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":0,"alsoIn":[],"x":0.6854,"y":0.48,"text":"OpenAI presents a method that uses a large language model (GPT-4) to automatically generate natural-language explanations of the behavior of individual neurons in a smaller subject language model (GPT-2 XL). Each explanation is produced from examples of text tokens paired with the neuron's activations, then quantitatively scored by having a simulator model predict the neuron's activations from the explanation and correlating those predictions against the neuron's real activations. The work studies how explanation quality scales with explainer/simulator/subject model size and training time, introduces an explanation-revision loop and 'neuron puzzles', and surfaces qualitative interpretable neurons. A recurring finding noted in the work is that GPT-4's generated explanations tend to be overly broad relative to the neuron's true behavior."},{"id":"arxiv:2306.05685","kind":"paper","n":293,"label":"Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena","authors":"Zheng et al.","year":2023,"url":"https://arxiv.org/abs/2306.05685","region":"t33","hemi":"llm","lobe":"parietal","color":"#ad67b0","indegree":3,"alsoIn":[],"x":0.8843,"y":0.6594,"text":"Traditional LLM benchmarks (MMLU, HELM) fail to capture human preference for aligned chat assistants on open-ended tasks, so the paper proposes 'LLM-as-a-judge'\u2014using strong models like GPT-4 to evaluate chatbot responses\u2014and introduces two human-preference benchmarks, MT-bench (80 multi-turn questions across 8 categories) and Chatbot Arena (a crowdsourced anonymous-battle platform with ~30K votes). It systematically studies judge limitations (position, verbosity, self-enhancement biases and weak math/reasoning grading) and proposes mitigations (position swapping, few-shot, chain-of-thought, reference-guided grading). GPT-4 judges agree with human experts at over 80% (85% on non-tie MT-bench votes), matching the human-human agreement level of 81%. The work argues for a hybrid framework combining capability benchmarks with preference benchmarks, and releases MT-bench questions, 3K expert votes, and 30K human-preference conversations."},{"id":"arxiv:2407.12036","kind":"paper","n":294,"label":"Exploring Advanced Large Language Models with LLMsuite","authors":"","year":null,"url":"https://arxiv.org/abs/2407.12036","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":7,"alsoIn":[],"x":0.7901,"y":0.7592,"text":"This tutorial/survey paper reviews techniques for extending and improving Large Language Models (e.g., ChatGPT, Gemini) beyond their core, addressing limitations like knowledge cutoffs, arithmetic errors, and hallucinations. It surveys augmentation frameworks (Retrieval-Augmented Generation, Program-Aided Language Models, ReAct, LangChain), transformer architectures, distributed-training methods (DDP, FSDP, ZeRO, 1-bit BitNet b1.58), parameter-efficient fine-tuning (LoRA, prompt tuning), and human-alignment methods (RLHF, ReST, PPO). It is accompanied by a code toolbox/tutorial named LLMsuite (source code available on request) and a reference table of NLP datasets/benchmarks. The paper is pedagogical, consolidating existing methods rather than presenting new experimental results."},{"id":"arxiv:2305.14992","kind":"paper","n":295,"label":"Reasoning with Language Model is Planning with World Model","authors":"Hao et al.","year":2023,"url":"https://arxiv.org/abs/2305.14992","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":15,"alsoIn":[],"x":0.7651,"y":0.793,"text":"LLMs reason autoregressively and lack an internal world model to track environment/intermediate states, simulate action outcomes, and balance exploration vs. exploitation, causing failures on tasks easy for humans (e.g., GPT-3 plans Blocksworld at 1% vs 78% human). The paper proposes Reasoning via Planning (RAP), which repurposes a single LLM as both a reasoning agent (proposing actions) and a world model (predicting next states), and uses Monte Carlo Tree Search guided by reward functions to explore a reasoning tree. Across plan generation (Blocksworld), math reasoning (GSM8K), and logical inference (PrOntoQA), RAP consistently beats CoT, least-to-most, and self-consistency baselines. Notably, LLaMA-33B with RAP surpasses GPT-4 with CoT by 33% relative on Blocksworld plan generation."},{"id":"arxiv:2407.10311","kind":"paper","n":296,"label":"Sora and V-JEPA Have Not Learned The Complete Real World Model \u2014 A Philosophical Analysis of Video AIs Through the Theory of Productive Imagination","authors":"Jianqiu Zhang","year":2024,"url":"https://arxiv.org/abs/2407.10311","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":2,"alsoIn":[],"x":0.3663,"y":0.6238,"text":"The paper applies Kantian philosophy (the theory of productive imagination from the Critique of Pure Reason) to argue that neither OpenAI's Sora nor Meta's V-JEPA has learned a complete 'coherent world model.' It defines three indispensable components of a coherent world model \u2014 representations of isolated objects, an a priori law of change across space and time, and Kantian categories \u2014 and diagnoses each system against them. It concludes Sora operates only in a 'dreaming mode' (lacking the a priori law of change and Kantian categories, flaws not fixable by scaling), while V-JEPA captures only the context-dependent aspect of the law of change but misses Kantian categories and experience. It then proposes a Coherent World Model Learning Architecture (CWMLA) that aligns encoder outputs from fragmented/out-of-order clips with those from ordered clips to build a genuine world model."},{"id":"arxiv:2411.04983","kind":"paper","n":297,"label":"DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning","authors":"Gaoyue Zhou, Hengkai Pan, Yann LeCun, Lerrel Pinto","year":2024,"url":"https://arxiv.org/abs/2411.04983","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":1,"alsoIn":[],"x":0.2522,"y":0.6546,"text":"World models (action-conditioned predictive models) are usually task-specific, trained online, and require auxiliary information like rewards, demos, or inverse models, limiting their generality. DINO-WM trains a task-agnostic world model on offline trajectories by predicting future DINOv2 patch features with a causal ViT transition model, then solves novel tasks at test time as visual goal-reaching via model predictive control (CEM) over actions in latent space \u2014 without expert demos, reward models, or inverse models. Across six environments (maze navigation, Push-T, Reacher, rope and granular deformable manipulation), it achieves zero-shot planning that beats prior SOTA, improving success on the hardest tasks by ~45% on average and LPIPS prediction quality by ~56%. It also generalizes to unseen environment configurations (random walls/doors, novel object shapes, varying particle counts)."},{"id":"arxiv:2606.06556","kind":"paper","n":298,"label":"Robots Need More than VLA and World Models","authors":"(see arXiv)","year":2026,"url":"https://arxiv.org/abs/2606.06556","region":"t0","hemi":"llm","lobe":"frontal","color":"#9a4f4f","indegree":8,"alsoIn":[],"x":0.6226,"y":0.2863,"text":"This position paper argues that generalist robotics is bottlenecked not only by policy scaling (more demonstrations, larger VLA models) but by the absence of mechanisms that convert the world's abundant unstructured behavioural data\u2014human motion, internet video, simulation rollouts, deployment traces\u2014into grounded robot supervision with action labels, task semantics, and reward structure. It surveys robot-native datasets and VLAs, learning from weakly-grounded video, simulation/world-model experience generation, and reward modelling, organising the literature by the supervision bottleneck each exposes. It identifies four missing components for a 'grounding-centric' pipeline: a physical data engine for embodied autolabelling, task-preserving retargeting across embodiments, physics-grounded world models for consequence prediction, and a self-improving deployment loop with task-conditioned reward grounding. The central claim is that the next robotics foundation model will be a compounding pipeline that grounds physical experience into actions, rewards, world models, and deployment feedback\u2014not a single VLA or world model."},{"id":"arxiv:2603.13910","kind":"paper","n":299,"label":"Scene Generation at Absolute Scale: Semantic and Geometric Guidance for 3D Indoor Scenes","authors":"n/a","year":2026,"url":"https://arxiv.org/abs/2603.13910","region":"t15","hemi":"llm","lobe":"occipital","color":"#4f9a62","indegree":2,"alsoIn":[],"x":0.5732,"y":0.1288,"text":"GuidedSceneGen is a text-to-3D indoor scene generation framework that maintains an absolute metric world coordinate frame throughout the entire pipeline, addressing geometric drift and scale ambiguity in prior text-to-3D methods. From a text prompt it first predicts a global 3D layout (via Holodeck) encoding semantics and geometry, then synthesizes a 360\u00b0 cubemap panorama with a semantics- and depth-conditioned multi-view diffusion model, explores unobserved regions with a camera-guided video diffusion model using collision-aware optimized trajectories, and fuses views with 3D Gaussian Splatting. It achieves the highest CLIP score (32.48) and best multi-view consistency (MEt3R 0.024) among baselines, is preferred in a user study, runs ~10x faster than WorldExplorer (0.75h vs 7h), and enables accurate transfer of 9D poses/semantic labels plus seamless scene expansion without re-alignment."},{"id":"arxiv:2509.02547","kind":"paper","n":300,"label":"The Landscape of Agentic Reinforcement Learning for LLMs: A Survey","authors":"Guibin Zhang et al.","year":2025,"url":"https://arxiv.org/abs/2509.02547","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":9,"alsoIn":[],"x":0.7123,"y":0.8722,"text":"This survey formalizes the shift from conventional preference-based RL fine-tuning of LLMs (modeled as a degenerate single-step MDP with horizon T=1) to Agentic RL, where LLMs are cast as policies in temporally extended, partially observable MDPs (POMDPs) embedded in dynamic environments. It proposes a twofold taxonomy\u2014one organized around core agentic capabilities (planning, tool use, memory, reasoning, self-improvement, perception) and one around task domains (search, code, math, GUI, vision, embodied, multi-agent)\u2014and argues RL is the critical mechanism that converts these from static heuristic modules into adaptive behavior. Synthesizing over 500 works, it also consolidates open-source environments, benchmarks, and RL frameworks into a practical compendium and outlines open challenges in trustworthiness, scaling training, and scaling environments."},{"id":"arxiv:2501.02189","kind":"paper","n":301,"label":"A Survey of State of the Art Large Vision Language Models: Alignment, Benchmark, Evaluations and Challenges","authors":"Zongxia Li, Xiyang Wu, Hongyang Du, Fuxiao Liu, Huy Nghiem, Guangyao Shi","year":2025,"url":"https://arxiv.org/abs/2501.02189","region":"t30","hemi":"llm","lobe":"temporal","color":"#8c67b0","indegree":8,"alsoIn":[],"x":0.7225,"y":0.2288,"text":"This survey reviews state-of-the-art multimodal Vision-Language Models (VLMs) developed up to 2025, covering their building-block architectures, training/alignment methods, benchmarks and evaluation metrics, and open challenges. It documents the architectural shift from training-from-scratch contrastive models (CLIP, BLIP, ALIGN) to using pretrained LLMs as backbones with projector/cross-attention fusion, catalogs ~30 SoTA VLMs (Table 1), and reviews alignment via RLHF, DPO, PPO, GRPO and RLVR. It collects and categorizes 95 benchmarks into 13 categories (Table 3) and critiques evaluation practices, noting most benchmarks use Yes/No or multiple-choice formats for ease of scoring. It closes by surveying challenges: hallucination, safety/jailbreaking, fairness, multimodal alignment, training efficiency, and data scarcity."},{"id":"arxiv:2509.13281","kind":"paper","n":302,"label":"RepIt: Steering Language Models with Concept-Specific Refusal Vectors","authors":"et al.","year":2025,"url":"https://arxiv.org/abs/2509.13281","region":"t8","hemi":"llm","lobe":"frontal","color":"#a2b067","indegree":2,"alsoIn":[],"x":0.7781,"y":0.6487,"text":"Current safety evaluations of language models rely on benchmark-based assessments that may miss localized vulnerabilities. RepIt is a three-step disentanglement procedure (reweighting, whitening, orthogonalization) that isolates concept-specific refusal vectors from collinear non-target concepts in activation space, enabling selective suppression of refusal on targeted harmful categories (e.g., WMD) while preserving refusal elsewhere. Across five frontier LMs, RepIt produces evaluation-evading model organisms that achieve high target-category jailbreak rates while maintaining near-baseline safety on standard benchmarks, exposing critical blind spots in current safety evaluation practices. The method is highly data-efficient, requiring as few as 12 target examples, and its edit localizes to just 100\u2013200 residual dimensions."},{"id":"arxiv:2511.20340","kind":"paper","n":303,"label":"Scaling LLM Speculative Decoding: Non-Autoregressive Forecasting in Large-Batch Scenarios","authors":"","year":null,"url":"https://arxiv.org/abs/2511.20340","region":"t9","hemi":"llm","lobe":"parietal","color":"#97b067","indegree":9,"alsoIn":[],"x":0.8414,"y":0.7696,"text":"Speculative decoding (SD) accelerates LLM inference by using idle compute during memory transfer, but mainstream batching already absorbs that idle compute, shrinking the usable draft-token budget and breaking SD methods that rely on large autoregressive draft trees. The paper proposes SpecFormer, a draft architecture that combines a unidirectional Context Causal Attention (extracting features from the full input via multi-layer hidden-state fusion, a downsampler, causal attention, and a Positional FFN) with a bidirectional Draft Bi-directional Attention that generates all draft tokens in parallel, making most parameters position-independent. It introduces an evaluation framework based on arithmetic intensity, a redundancy ratio \u03c1, and an optimization coefficient \u03ba=a\u00b7l_d/k to measure SD efficiency under a fixed draft-token budget. Across Qwen and LLaMA models (4B/8B/14B) it delivers consistent lossless speedups (~1.78\u20131.81\u00d7 in Table 1) that persist even at large batch sizes up to 128, outperforming HASS and EAGLE-3 under constrained draft budgets."},{"id":"github.com:14ee211b0b5d","kind":"paper","n":304,"label":"NVIDIA/Cosmos \u2014 omnimodal world-model platform for Physical AI, served three ways","authors":"NVIDIA (GitHub)","year":2025,"url":"https://github.com/NVIDIA/Cosmos","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":5,"alsoIn":[],"x":0.4315,"y":0.9533,"text":"NVIDIA Cosmos 3 is an open platform of omnimodal world foundation models that jointly process and generate language, images, video, audio, and action sequences within a unified Mixture-of-Transformers (MoT) architecture. It exposes two runtime surfaces \u2014 a Reasoner (autoregressive transformer producing text for world understanding, physical reasoning, planning, and embodied agent reasoning) and a Generator (diffusion transformer producing vision/sound/action for world simulation, future prediction, and synthetic data). The family spans a 16B Nano and 64B Super model plus specialized variants for text-to-image, image-to-video, and DROID robot policy, targeting Physical AI for robots, autonomous vehicles, and smart infrastructure."},{"id":"people.idsia.ch:73d35fac50f9","kind":"paper","n":305,"label":"Making the World Differentiable: On Using Self-Supervised Fully Recurrent Neural Networks for Dynamic Reinforcement Learning and Planning in Non-Stationary Environments","authors":"Schmidhuber","year":1990,"url":"https://people.idsia.ch/~juergen/FKI-126-90_(revised)bw_ocr.pdf","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":1,"alsoIn":[],"x":0.0785,"y":0.3495,"text":"The paper attacks the temporal credit-assignment problem in reinforcement learning within non-stationary, non-Markovian reactive environments where arbitrary time lags separate actions from consequences. Its core method couples two interacting fully recurrent networks: a controller and a self-supervised 'model network' that predicts ALL of the controller's future inputs (including scalar reinforcement / 'pain' units), thereby 'making the world differentiable' so that ordinary supervised gradient descent (via the forward-in-time IID/RTRL algorithm) can propagate error through a learned model of the environment back to the controller. It presents sequential and parallel on-line versions, plus extensions: a multi-dimensional recurrent adaptive critic (TD), planning by 'mental simulation' of the model forward in time, and dynamic curiosity/boredom that rewards actions increasing the model's predictive knowledge. Proof-of-concept experiments solve a non-Markovian reinforcement flip-flop and improve non-Markovian cart-pole balancing."},{"id":"arxiv:2506.21521","kind":"paper","n":306,"label":"Potemkin Understanding in Large Language Models","authors":"Mancoridis, Weeks, Vafa & Mullainathan","year":2025,"url":"https://arxiv.org/abs/2506.21521","region":"t33","hemi":"llm","lobe":"parietal","color":"#ad67b0","indegree":2,"alsoIn":[],"x":0.7901,"y":0.7287,"text":"The paper asks what justifies inferring an LLM's conceptual understanding from its benchmark answers, arguing that human-designed benchmarks (AP exams, AIME, etc.) are only valid for LLMs if LLM misunderstandings are structured like human ones. It formalizes 'keystone' question sets and defines 'potemkin understanding' as the case where a model answers keystones correctly yet fails to apply the concept\u2014an answer pattern no human would produce. Using a purpose-built benchmark (32 concepts across literary techniques, game theory, and psychological biases) and a separate automated lower-bound procedure, it shows potemkins are ubiquitous: models define concepts ~94% correctly but fail to use them, with high potemkin rates across all 7+ models. It further shows these failures stem from internal incoherence\u2014models often disagree with their own generated examples."},{"id":"arxiv:2411.08794","kind":"paper","n":307,"label":"LLM-Based World Models Can Make Decisions Solely, But Rigorous Evaluations are Needed","authors":"Chang Yang, Xinrun Wang, Junzhe Jiang, Qinggang Zhang, Xiao Huang","year":2024,"url":"https://arxiv.org/abs/2411.08794","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":18,"alsoIn":[],"x":0.7093,"y":0.8852,"text":"The paper argues that LLM-based world models can make decisions on their own (not just predict transitions to support a separate agent/planner), but that current evaluation practices\u2014treating them as general world simulators or as coupled functional modules\u2014are inadequate. It presents two observations motivating sole decision-making and three observations exposing flaws in current evaluation, then proposes a decoupled bottom-up evaluation framework of three tasks: policy verification, action proposal, and policy planning. Using GPT-4o and GPT-4o-mini across 31 text-based environments (from ByteSized32/Wang et al.) with curated rule-based ground-truth policies (verified over 200 runs, 30 trials per setting, temperature 0), the authors find GPT-4o consistently beats GPT-4o-mini\u2014especially on domain-knowledge-heavy scientific tasks\u2014that performance hinges on critical bottleneck steps rather than total step count, and that combining functionalities increases performance instability."},{"id":"arxiv:2502.14739","kind":"paper","n":308,"label":"SuperGPQA: Scaling LLM Evaluation across 285 Graduate Disciplines","authors":"M-A-P Team et al.","year":2025,"url":"https://arxiv.org/abs/2502.14739","region":"t33","hemi":"llm","lobe":"parietal","color":"#ad67b0","indegree":6,"alsoIn":[],"x":0.8155,"y":0.5946,"text":"SuperGPQA is a graduate-level benchmark of 26,529 multiple-choice questions spanning 285 disciplines (13 disciplines, 72 fields, 285 subfields), built to evaluate LLM knowledge/reasoning in long-tail fields (e.g. light industry, agriculture, service disciplines) underserved by prior benchmarks. It uses a Human-LLM collaborative filtering pipeline (source screening \u2192 transcription \u2192 3-stage quality inspection) with 80+ expert annotators to remove trivial/ambiguous items, yielding an average of 9.67 options per question. Evaluating 6 reasoning, 28 chat, and 17 base models, the best model DeepSeek-R1 reaches only 61.82% accuracy, exposing a large gap to AGI. Analysis shows reasoning capacity and instruction tuning drive performance, humanities disciplines discriminate models best, and STEM (chemistry/physics/math) is hardest across nearly all models."},{"id":"arxiv:2101.07393","kind":"paper","n":309,"label":"Grounding Language to Entities and Dynamics for Generalization in Reinforcement Learning","authors":"Hanjie, Zhong & Narasimhan","year":2021,"url":"https://arxiv.org/abs/2101.07393","region":"t22","hemi":"wm","lobe":"temporal","color":"#4f849a","indegree":2,"alsoIn":[],"x":0.1537,"y":0.491,"text":"The paper tackles language grounding for generalization in reinforcement learning where, unlike prior work, no ground-truth mapping between text references and environment entity symbols is given \u2014 the agent must jointly ground a crowdsourced game manual to entity symbols and movement dynamics using only scalar rewards. It introduces MESSENGER, a multi-task grid-world environment with free-form English manuals, and EMMA (Entity Mapper with Multi-modal Attention), an end-to-end differentiable model that uses entity-conditioned attention over BERT-encoded descriptions to build text-conditioned entity representations for the control policy. EMMA achieves strong zero-shot generalization to unseen games (>40% higher win rate than baselines), but all models including EMMA remain near 10% win rate on the hardest stage (S3) that requires grounding language to movement dynamics, far below the 84% human level."},{"id":"arxiv:2304.03442","kind":"paper","n":310,"label":"Generative Agents: Interactive Simulacra of Human Behavior","authors":"Park et al.","year":2023,"url":"https://arxiv.org/abs/2304.03442","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":1,"alsoIn":[],"x":0.5334,"y":0.9398,"text":"The paper introduces generative agents: LLM-backed software agents that simulate believable individual and emergent social human behavior. The core contribution is an architecture extending an LLM (gpt-3.5-turbo/ChatGPT) with a natural-language memory stream plus a retrieval function (scoring memories by recency, relevance, and importance), periodic reflection that synthesizes observations into higher-level inferences, and recursive top-down planning that keeps behavior coherent over time. Twenty-five agents are instantiated in a Sims-style sandbox town (Smallville), where a controlled 'interview' evaluation and a two-day end-to-end simulation show the full architecture is the most believable and that emergent behaviors (information diffusion, relationship formation, party coordination) arise from a single seed."},{"id":"arxiv:2412.07077","kind":"paper","n":311,"label":"Retaining and Enhancing Pre-trained Knowledge in Vision-Language Models with Prompt Ensembling","authors":"","year":null,"url":"https://arxiv.org/abs/2412.07077","region":"t23","hemi":"wm","lobe":"occipital","color":"#4f799a","indegree":4,"alsoIn":[],"x":0.3282,"y":0.3743,"text":"The paper tackles the problem that CLIP loses its zero-shot generalization when fine-tuned on specialized or fine-grained datasets. It proposes Group-wise Prompt Ensemble (GPE), which divides learnable prompts into two main groups plus auxiliary prompts, uses masked (read-only) attention to prevent representation shifts, trains groups separately with covariance regularization to diversify prompts, and fuses all prompts plus CLIP's special token at inference. GPE achieves the best average harmonic mean (79.24) across 11 datasets in base-to-new generalization and the highest average cross-dataset transfer (63.17), retaining near zero-shot performance even after fine-tuning on niche domains."},{"id":"arxiv:2503.15558","kind":"paper","n":312,"label":"Cosmos-Reason1: From Physical Common Sense To Embodied Reasoning","authors":"NVIDIA (Alisson Azzolini et al.)","year":2025,"url":"https://arxiv.org/abs/2503.15558","region":"t7","hemi":"llm","lobe":"occipital","color":"#adb067","indegree":1,"alsoIn":[],"x":0.6764,"y":0.3228,"text":"Cosmos-Reason1 introduces two multimodal LLMs (7B dense and 56B hybrid Mamba-MLP-Transformer) that perceive the physical world via video input and produce embodied decisions through long chain-of-thought reasoning. The authors define a hierarchical physical-common-sense ontology (3 categories, 16 subcategories) and a 2D embodied-reasoning ontology, curate ~4M video-text annotations, and train in two stages: Physical AI SFT and Physical AI RL with rule-based verifiable rewards (including self-supervised MCQs like spatial puzzles and arrow-of-time). They build new benchmarks for physical common sense (604 questions/426 videos) and embodied reasoning (610 questions/600 videos). SFT lifts backbone VLMs >10% on embodied reasoning and RL adds a further ~5%, with large gains on intuitive physics where prior models perform near chance."},{"id":"pubmed.ncbi.nlm.nih.gov:fd66789ead66","kind":"paper","n":313,"label":"The hippocampus as a spatial map: preliminary evidence from unit activity in the freely-moving rat","authors":"O'Keefe, J. & Dostrovsky, J.","year":1971,"url":"https://pubmed.ncbi.nlm.nih.gov/5124915/","region":"t28","hemi":"llm","lobe":"frontal","color":"#7667b0","indegree":3,"alsoIn":[],"x":0.5583,"y":0.4167,"text":"This is the seminal report by O'Keefe and Dostrovsky recording unit (single-neuron) activity from the hippocampus of freely-moving rats. It presents preliminary electrophysiological evidence that some hippocampal neurons fire selectively as a function of the animal's spatial position and orientation in its environment. The authors argue these findings support the idea that the hippocampus functions as a spatial map (cognitive map) of the environment. No abstract or quantitative results are available in the provided record."},{"id":"arxiv:2502.10498","kind":"paper","n":314,"label":"The Role of World Models in Shaping Autonomous Driving: A Comprehensive Survey","authors":"Tu et al.","year":2025,"url":"https://arxiv.org/abs/2502.10498","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":8,"alsoIn":[],"x":0.3745,"y":0.772,"text":"This survey comprehensively reviews Driving World Models (DWMs), which predict the temporal evolution of driving scenes conditioned on multimodal instructions to support autonomous driving. It organizes the field along three axes: the DWM ecosystem (simulators, high-impact datasets, and multi-dimensional metrics), a taxonomy by prediction modality (video/visual, point cloud and occupancy in 4D space, multimodal, latent, and vectorized/traffic-map spaces), and applications (simulation, data generation, enhancing driving/planning, and pre-training). It compiles performance tables for representative methods across video, point cloud, and occupancy generation plus open- and closed-loop planning, then outlines limitations (data scarcity, reliable simulation, task unification, multi-sensor modeling, efficiency, adversarial attack/defense) and future directions."},{"id":"nature.com:ad56347f1921","kind":"paper","n":315,"label":"The hippocampus as a predictive map","authors":"Stachenfeld, K. L., Botvinick, M. M. & Gershman, S. J.","year":2017,"url":"https://www.nature.com/articles/nn.4650","region":"t28","hemi":"llm","lobe":"frontal","color":"#7667b0","indegree":7,"alsoIn":[],"x":0.5628,"y":0.5299,"text":"The paper asks, from a reinforcement-learning perspective, what spatial representation best supports maximizing future reward, and argues the answer is a predictive representation \u2014 the successor representation (SR), which encodes expected discounted future state occupancy under the current policy. Modeling hippocampal place cells as encoding the SR reproduces non-spatial properties of place fields (predictive skewing, reward/goal sensitivity, policy dependence) that a purely geometric cognitive map cannot explain. The authors further propose that entorhinal grid cells encode the eigenvectors of the SR, providing a low-dimensional basis set that denoises predictions and exposes multiscale structure for hierarchical planning. The model is shown to match empirical place- and grid-cell phenomena across spatial, nonspatial, spatiotemporal, geometric, and multicompartment environments (Figs 1\u20138)."},{"id":"arxiv:2405.17398","kind":"paper","n":316,"label":"Vista: A Generalizable Driving World Model with High Fidelity and Versatile Controllability","authors":"Gao et al. (OpenDriveLab)","year":2024,"url":"https://arxiv.org/abs/2405.17398","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":4,"alsoIn":[],"x":0.2963,"y":0.7423,"text":"Vista is a generalizable driving world model built on Stable Video Diffusion that predicts high-fidelity future driving scenes at 10 Hz and 576\u00d71024 resolution. It introduces a latent-replacement scheme to inject position/velocity/acceleration priors for coherent long-horizon rollouts, two novel losses (dynamics enhancement and structure preservation) to improve fidelity, and a unified interface for versatile multi-modal action controls (command, goal point, trajectory, angle, speed) learned via LoRA and collaborative training across OpenDV-YouTube and nuScenes. It outperforms the best prior driving world model by 55% in FID (6.9 vs 15.4) and 27% in FVD (89.4 vs 122.7) on nuScenes, and beats general-purpose video generators in over 70% of human-evaluation comparisons. Vista's own prediction uncertainty is also repurposed as a generalizable, ground-truth-free reward function for action evaluation."},{"id":"arxiv:2507.21141","kind":"paper","n":317,"label":"Death by a Thousand Directions: Exploring the Geometry of Harmfulness in LLMs through Subconcept Probing","authors":"et al.","year":2025,"url":"https://arxiv.org/abs/2507.21141","region":"t8","hemi":"llm","lobe":"frontal","color":"#a2b067","indegree":3,"alsoIn":[],"x":0.743,"y":0.6807,"text":"The paper introduces a multidimensional framework for probing and steering harmful content in LLM internals by training one linear probe for each of 55 harmfulness subconcepts (e.g., racial hate, employment scams, weapons) on the attention-output hidden states of Llama-3.1-8B-Instruct (replicated on Qwen-2-7B-Instruct). Stacking the 55 probe weight vectors and applying SVD reveals that the resulting harmfulness subspace is strikingly low-rank: at energy \u03c4=0.95 nearly every layer collapses to effective rank K=1. The authors then ablate the full subspace, and ablate/steer along the single dominant direction; dominant-direction steering nearly eliminates jailbreak success with only minor utility loss. They conclude that concept subspaces are a scalable, low-rank lens for auditing and hardening LLM safety behaviour."},{"id":"arxiv:2511.20644","kind":"paper","n":318,"label":"Vision-Language Memory for Spatial Reasoning (VLM2)","authors":"Anonymous et al.","year":2025,"url":"https://arxiv.org/abs/2511.20644","region":"t12","hemi":"llm","lobe":"frontal","color":"#5e9a4f","indegree":0,"alsoIn":[],"x":0.5972,"y":0.1318,"text":"VLM2 is a video-only vision-language model that tackles two failures in video-based spatial reasoning: semantic-geometric feature misalignment and lack of persistent memory. It builds a view-consistent 3D-aware representation from 2D video via adaptive 3D position injection, viewpoint-aware geometry alignment, and cross-attention semantic-geometric fusion, then adds a dual-memory module (sliding-window working memory + fixed-capacity episodic memory) for bounded long-horizon reasoning. Built on LLaVA-Video-7B with \u03c03 as the 3D foundation model, it sets state-of-the-art among video-only models, e.g. 68.8 avg on VSI-Bench (+7.9 over VLM-3R) and 65.3 on VSTI-Bench (+6.5)."},{"id":"arxiv:2310.04276","kind":"paper","n":319,"label":"From Task Structures to World Models: What Do LLMs Know?","authors":"Yildirim & Paul","year":2024,"url":"https://arxiv.org/abs/2310.04276","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":3,"alsoIn":[],"x":0.7228,"y":0.7435,"text":"This opinion/theory article asks in what sense large language models like GPT-4 'know' anything. The authors grant LLMs 'instrumental knowledge'\u2014knowledge defined by the ability to infer task structure from context and condition next-token prediction on it\u2014and contrast this with humans' 'worldly knowledge', which is grounded in structured world models (homomorphic, causal abstractions of objects, scenes, and agents from cognitive science). Reviewing probing studies (Othello-GPT, program synthesis, protein folding, perceptual color space), they argue next-token prediction/compression can partially recover an underlying data-generating process, but that this recovery is weak for complex real-world structure. They propose that the degree to which LLMs acquire worldly knowledge is governed by an implicit, resource-rational tradeoff between costly world models and frequently-occurring task demands, and that explicit world models offer a path to safer, more interpretable, better-aligned AI."},{"id":"arxiv:2511.08585","kind":"paper","n":320,"label":"Simulating the Visual World with Artificial Intelligence: A Roadmap","authors":"(multi-author)","year":2025,"url":"https://arxiv.org/abs/2511.08585","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":13,"alsoIn":[],"x":0.3896,"y":0.8112,"text":"This survey proposes a roadmap for how video generation models are evolving into physical world models, conceptualizing a modern video foundation model as the combination of an implicit world model (a latent simulation engine encoding physical laws, interaction dynamics, and agent behavior) and a video renderer (which turns latent states into observable video). It introduces a four-generation taxonomy organized around three core capabilities\u2014faithfulness, interactiveness, and planning\u2014tracing progress from short photorealistic clips (Gen 1) to controllable interaction (Gen 2), real-time complex planning (Gen 3), and stochastic, multi-scale modeling of rare events (Gen 4). It formalizes the world model as next-scene video prediction V_{1:T}=G(I), introduces the concept of a 'navigation mode' distinguished from spatial conditions, and catalogs representative methods across robotics, autonomous driving, gaming, and general scenes. The paper closes with open challenges and design principles for next-generation, 'everything-everywhere-anytime' world models."},{"id":"arxiv:2205.11916","kind":"paper","n":321,"label":"Large Language Models are Zero-Shot Reasoners","authors":"Takeshi Kojima et al.","year":2022,"url":"https://arxiv.org/abs/2205.11916","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":7,"alsoIn":[],"x":0.8828,"y":0.7912,"text":"The paper challenges the view that large language models only reason well via few-shot exemplars by showing they are competent zero-shot reasoners when simply prompted with 'Let's think step by step.' The proposed Zero-shot-CoT uses a single task-agnostic trigger in a two-stage pipeline (reasoning extraction, then answer extraction) and is evaluated across 12 arithmetic, commonsense, symbolic, and logical reasoning benchmarks with 17 models. It dramatically lifts zero-shot accuracy on multi-step tasks (e.g., MultiArith 17.7%\u219278.7%, GSM8K 10.4%\u219240.7% with text-davinci-002) and replicates on 540B PaLM. The single prompt generalizes across diverse tasks, suggesting untapped high-level zero-shot cognitive capabilities hidden in LLMs."},{"id":"arxiv:2605.09131","kind":"paper","n":322,"label":"MCP-Cosmos: World Model-Augmented Agents for Complex Task Execution in MCP Environments","authors":"Ganapavarapu & Patel","year":2026,"url":"https://arxiv.org/abs/2605.09131","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":1,"alsoIn":[],"x":0.5664,"y":0.9344,"text":"MCP-Cosmos addresses the gap between task-level planning (which ignores execution dynamics) and reactive execution (which lacks long-horizon foresight) by infusing generative World Models into the Model Context Protocol agent loop, letting agents simulate tool-call state transitions in latent space before real execution. The framework uses a 'Bring Your Own World Model' (BYOWM) design and is evaluated with two proactive strategies (ReAct-Plan-Exec and SPIRAL-Exec) against a ReAct baseline, using 2 planners (gpt-oss-120b, claude-sonnet-4.6) and 3 world models (gpt-oss-120b, claude-sonnet-4.6, Arctic-AWM-4B) over a 24-task subset of MCP-Bench (300+ trajectories). World-model-augmented agents improve tool selection, parameter accuracy and execution efficiency over the baseline, and the authors propose a new Execution Quality metric that penalizes excessive/failed tool calls. They find purpose-built small MCP world models (AWM-4B) underperform general-purpose LLM world models, and that stronger planners benefit more from explicit world models on the efficiency axis."},{"id":"arxiv:2406.18043","kind":"paper","n":323,"label":"GenRL: Multimodal-foundation world models for generalization in embodied agents","authors":"Mazzaglia, Verbelen, Dhoedt, Courville & Rajeswar","year":2024,"url":"https://arxiv.org/abs/2406.18043","region":"t0","hemi":"wm","lobe":"frontal","color":"#9a4f4f","indegree":1,"alsoIn":[],"x":0.1952,"y":0.6739,"text":"GenRL tackles building generalist embodied agents that can solve many tasks specified by language or visual prompts without per-task reward engineering or any language annotations in the embodied domain. It learns multimodal-foundation world models (MFWMs) that connect and align a frozen video-language foundation model (InternVideo2) with the discrete latent space of a DreamerV3-style generative world model using vision-only data, then turns prompts into latent targets and learns matching behaviors purely in imagination via a trajectory-matching RL reward. Across 35 reward-free tasks in 4 locomotion environments and a kitchen manipulation environment, GenRL outperforms image/video-language reward baselines and a reversed-connector world-model baseline. It further introduces data-free policy learning, generalizing to new tasks after pre-training with no access to any data."},{"id":"arxiv:2303.00915","kind":"paper","n":324,"label":"BiomedCLIP: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs","authors":"Sheng Zhang et al.","year":2023,"url":"https://arxiv.org/abs/2303.00915","region":"t13","hemi":"llm","lobe":"parietal","color":"#6bb067","indegree":4,"alsoIn":[],"x":0.6585,"y":0.1135,"text":"Existing biomedical vision-language models are limited by small, private, and undiverse pretraining data mostly focused on chest X-rays. The authors build PMC-15M, a public dataset of 15,282,336 image-caption pairs from 4.4 million PubMed Central articles spanning ~30 image types (two orders of magnitude larger than MIMIC-CXR), and pretrain BiomedCLIP, a CLIP adaptation using a PubMedBERT text encoder, WordPiece tokenizer, 256-token context, and a ViT-B/16 image encoder. BiomedCLIP sets new state of the art across cross-modal retrieval, zero-shot/linear-probe image classification, and medical VQA, and even surpasses radiology-specific models like BioViL on RSNA pneumonia detection. It also enables privacy-preserving analysis by retrieving PMC-15M proxies for proprietary patient data."},{"id":"philpapers.org:9a9c4dcb81d9","kind":"paper","n":325,"label":"Analytical Biology","authors":"Gerd Sommerhoff","year":1950,"url":"https://philpapers.org/rec/SOMAB-4","region":"t2","hemi":"llm","lobe":"temporal","color":"#b07d67","indegree":4,"alsoIn":[],"x":0.6181,"y":0.7323,"text":"No paper content is available. The provided text is a Cloudflare security block page from philpapers.org, not the actual paper titled 'Analytical Biology'. No abstract, methods, results, or figures could be read, so no substantive claims can be indexed."},{"id":"arxiv:1701.06970","kind":"paper","n":326,"label":"A framework for cascade size calculations on random networks","authors":"","year":null,"url":"https://arxiv.org/abs/1701.06970","region":"t36","hemi":"llm","lobe":"frontal","color":"#b06793","indegree":3,"alsoIn":[],"x":0.6316,"y":0.7428,"text":"The paper presents an exact analytic framework to compute the full time evolution of average cascade size for a broad class of cascade models on random network ensembles (configuration-model graphs with arbitrary degree distribution, degree-degree correlations, and arbitrary threshold distributions) in the infinite-network-size limit. Instead of the standard generating-function/branching-process approximation, it reformulates the local tree approximation (LTA) as an iterative update of probability distributions over node load, decomposing a node's load into a sum of independent random variables convolved via FFT. This lets it handle history-dependent, continuous load redistribution that branching-process methods cannot, demonstrated on a fiber bundle model that was previously analytically intractable, plus constant-load models extended to degree-degree correlations. Numerical LTA shows perfect agreement with Monte Carlo simulations while reducing runtime from hours to under a minute."},{"id":"arxiv:2411.02385","kind":"paper","n":327,"label":"How Far is Video Generation from World Model: A Physical Law Perspective","authors":"Kang, Yue, Lu et al.","year":2024,"url":"https://arxiv.org/abs/2411.02385","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":5,"alsoIn":[],"x":0.4362,"y":0.838,"text":"The paper investigates whether scaling diffusion-based video generation models lets them discover fundamental physical laws purely from vision. Using a Box2D 2D simulator of object motion and collisions (uniform motion, elastic collision, parabolic motion) plus the PHYRE testbed for combinatorial settings, the authors scale data (30K\u20136M videos) and DiT models (22M\u2013456M params) and quantitatively measure velocity-prediction error across in-distribution (ID), out-of-distribution (OOD), and combinatorial generalization. They find near-perfect ID generalization, measurable scaling benefit for combinatorial generalization, but failure on OOD that scaling does not fix. Deeper analysis shows the models generalize 'case-based' (mimicking the nearest training example) and prioritize attributes in the order color > size > velocity > shape, concluding scaling alone is insufficient to learn physical laws."},{"id":"blogs.nvidia.com:9f8a6e618fbd","kind":"paper","n":328,"label":"Cosmos World Foundation Models Now Openly Available to Physical AI Developer Community","authors":"NVIDIA","year":2025,"url":"https://blogs.nvidia.com/blog/cosmos-world-foundation-models/","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":8,"alsoIn":[],"x":0.4405,"y":0.947,"text":"NVIDIA announced the Cosmos platform at CES 2025, releasing a family of open world foundation models (WFMs) \u2014 diffusion and autoregressive transformers that generate physics-aware videos of future environment states to accelerate physical AI development for robots and autonomous vehicles. The models were trained on 9,000 trillion tokens from 20 million hours of real-world video, come in Nano/Super/Ultra tiers with 4\u201314 billion parameters, and support text-to-world and video-to-world generation for synthetic data generation, simulation, and reinforcement learning. The release also includes video/image tokenizers, a NeMo Curator data pipeline, Cosmos Guardrails safety models, and a prompt-upsampling model, all under a permissive open model license."},{"id":"arxiv:2506.22355","kind":"paper","n":329,"label":"Embodied AI Agents: Modeling the World","authors":"Pascale Fung, Yoram Bachrach, Asli Celikyilmaz, Kamalika Chaudhuri, Delong Chen,","year":2025,"url":"https://arxiv.org/abs/2506.22355","region":"t7","hemi":"llm","lobe":"occipital","color":"#adb067","indegree":5,"alsoIn":[],"x":0.6331,"y":0.4204,"text":"This position/overview paper from Meta AI Research argues that world models are the central component for embodied AI agents (virtual avatars, wearable/AI-glasses agents, and robots), enabling reasoning, planning, and prediction beyond what generative next-token/next-pixel models offer. It frames world modeling as integrating multimodal perception, physical and mental (Theory-of-Mind) world models, action/control, and a scalable episodic memory, and contrasts inefficient generative models with efficient joint-embedding predictive (JEPA-style) approaches. It surveys Meta's concrete artifacts and benchmarks across the three agent types and reports large model-vs-human gaps on new physical/causal/procedural reasoning benchmarks. It closes with future directions on embodied learning (System A observation + System B action integration), multi-agent collaboration, and ethics (privacy, anthropomorphism)."},{"id":"arxiv:2402.10196","kind":"paper","n":330,"label":"A Trembling House of Cards? Mapping Adversarial Attacks against Language Agents","authors":"","year":null,"url":"https://arxiv.org/abs/2402.10196","region":"t8","hemi":"llm","lobe":"frontal","color":"#a2b067","indegree":3,"alsoIn":[],"x":0.814,"y":0.801,"text":"This position paper presents the first systematic mapping of adversarial attacks against LLM-powered language agents. It introduces a unified conceptual framework decomposing any agent into three components \u2014 Perception, Brain (reasoning/planning, working memory, long-term memory), and Action (tool use, embodiment) \u2014 and uses a hypothetical generalist agent ('Ultron') as a running example. Under this framework it enumerates 12 potential attack scenarios spanning input manipulation, jailbreaking/prompt injection, adversarial demonstrations, backdoors/data poisoning, malicious tools, and embodiment attacks, each linked to prior LLM attack literature. The work is a call to action: it argues there is a large gap between deployment speed and safety understanding, and urges studying agent risks before broad deployment, but reports no new experiments or quantitative results."},{"id":"arxiv:2304.14772","kind":"paper","n":331,"label":"Multisample Flow Matching: Straightening Flows with Minibatch Couplings","authors":"Aram-Alexandre Pooladian, Heli Ben-Hamu, Carles Domingo-Enrich, Brandon Amos, Ya","year":2023,"url":"https://arxiv.org/abs/2304.14772","region":"t3","hemi":"wm","lobe":"occipital","color":"#9a714f","indegree":5,"alsoIn":[],"x":0.3378,"y":0.1184,"text":"Existing simulation-free generative model training (diffusion, Flow Matching) constructs probability paths between independently sampled noise and data points, which leaves non-zero gradient variance at convergence and yields curved flows that are costly to simulate. The paper proposes Multisample Flow Matching, a generalization of Flow Matching that uses non-trivial minibatch couplings (e.g. batch optimal transport) between noise and data while exactly preserving the marginal constraints, all in a simulation-free, min-max-free manner. Theoretically, with BatchOT couplings the Joint CFM objective, flow straightness, and transport cost all converge to optimal as batch size k\u2192\u221e. Empirically on ImageNet-32/64 it cuts the number of function evaluations needed for a target FID by 30\u201360% with only ~4% extra training time and no loss in log-likelihood or sample quality."},{"id":"arxiv:2512.21577","kind":"paper","n":332,"label":"A Unified Definition of Hallucination, Or: It's the World Model, Stupid","authors":"","year":null,"url":"https://arxiv.org/abs/2512.21577","region":"t6","hemi":"llm","lobe":"temporal","color":"#b0a967","indegree":0,"alsoIn":[],"x":0.7198,"y":0.236,"text":"This position paper argues that the many fragmented, task-specific definitions of hallucination in LLMs can be unified under a single framework: hallucination is inaccurate internal world modeling that is observable to the user via outputs. It formalizes this with three components every evaluation implicitly assumes \u2014 a reference world model W (what is true), a view function V (what the model can observe), and a conflict resolution policy P with an associated truth function T that labels atomic claims true/false/unknown. By varying W, V, and P, the authors show prior definitions (summarization faithfulness, open-domain QA factuality, RAG, VLM, and agentic hallucination) are all special cases, and they argue this view enables fully-specified synthetic benchmarks (e.g., chess) where hallucination labels are determined by construction rather than human annotation. The paper connects to the companion HALLUWORLD benchmark and distinguishes true hallucination (wrong world beliefs) from planning errors and reward/incentive errors."},{"id":"arxiv:2603.20224","kind":"paper","n":333,"label":"Beyond Test-Time Compute Strategies: Advocating Energy-per-Token in LLM Inference","authors":"","year":null,"url":"https://arxiv.org/abs/2603.20224","region":"t24","hemi":"llm","lobe":"frontal","color":"#6784b0","indegree":5,"alsoIn":[],"x":0.8728,"y":0.7059,"text":"This position/analysis paper studies the energy-accuracy trade-off of test-time compute strategies (Chain-of-Thought prompting and Majority Voting) for small vs. large Llama models on the MMLU benchmark, measured on an NVIDIA L40S GPU. It shows that different transformer LLM architectures have distinct, nonlinear energy operating curves over input and generated token counts, and that CoT on a 1B model raises energy by 120-150x for accuracy gains that a larger 8B model achieves at far lower energy cost. The authors propose Energy-per-Token as a complementary efficiency metric and sketch an energy-aware query-routing architecture with performance and energy estimators that dynamically picks model and reasoning depth using operating curves."},{"id":"arxiv:2302.01560","kind":"paper","n":334,"label":"Describe, Explain, Plan and Select: Interactive Planning with Large Language Models Enables Open-World Multi-Task Agents","authors":"Wang, Cai, Liu, Ma & Liang","year":2023,"url":"https://arxiv.org/abs/2302.01560","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":2,"alsoIn":[],"x":0.5577,"y":0.9328,"text":"The paper tackles task planning for multi-task embodied agents in open-world environments like Minecraft, where long sub-goal dependency chains and state-dependent feasibility cause vanilla LLM planners to fail. It proposes DEPS (Describe, Explain, Plan and Select), an interactive LLM planning loop that describes execution outcomes, self-explains failures, re-plans, and uses a trainable horizon-predictive selector to rank parallel candidate sub-goals by estimated steps-to-completion. On 71 Minecraft tasks DEPS roughly doubles the average success rate over prior LLM planners (48.56% vs 25.77% for the best baseline) and is the first planning-based agent to obtain a diamond. It generalizes to ALFWorld (76% avg) and Tabletop manipulation (80% avg)."},{"id":"arxiv:2311.01017","kind":"paper","n":335,"label":"Copilot4D: Learning Unsupervised World Models for Autonomous Driving via Discrete Diffusion","authors":"Zhang et al. (Waabi & Univ. of Toronto)","year":2024,"url":"https://arxiv.org/abs/2311.01017","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":5,"alsoIn":[],"x":0.3455,"y":0.653,"text":"The paper addresses why unsupervised world models have scaled less successfully than GPT-style language models in robotics, identifying two bottlenecks: complex/unstructured observation spaces and the lack of a scalable parallel-decoding generative model. The authors propose Copilot4D, which tokenizes Lidar point cloud observations with a VQVAE-like tokenizer (using implicit volume rendering in the decoder), then predicts future frames via discrete diffusion using a spatio-temporal Transformer in Bird-Eye View. They recast MaskGIT as absorbing-uniform discrete diffusion with a few simple changes (injecting uniform noise during training and allowing iterative resampling of already-decoded tokens). Applied to point cloud forecasting, Copilot4D reduces prior SOTA Chamfer distance by 65-75% for 1s prediction and more than 50% for 3s prediction across NuScenes, KITTI Odometry, and Argoverse2."},{"id":"arxiv:1904.08149","kind":"paper","n":336,"label":"Bayesian policy selection using active inference","authors":"","year":null,"url":"https://arxiv.org/abs/1904.08149","region":"t14","hemi":"llm","lobe":"temporal","color":"#67b06e","indegree":4,"alsoIn":[],"x":0.5523,"y":0.9033,"text":"The paper addresses RL limitations such as reward shaping, poor generalization, and sample inefficiency by applying active inference and the free energy principle from neuroscience to artificial agents. Neural networks are used to learn a generative model comprising a state transition model, observation model, and likelihood model, casting action selection as inference by minimizing expected free energy. The approach is validated on the continuous mountain car problem with noisy observations, demonstrating that the learned dynamics model can predict environment evolution and that both expert-demonstration-based and reward-based preferred state priors yield successful policies."},{"id":"arxiv:2507.02076","kind":"paper","n":337,"label":"Reasoning on a Budget: A Survey of Adaptive and Controllable Test-Time Compute in LLMs","authors":"","year":null,"url":"https://arxiv.org/abs/2507.02076","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":4,"alsoIn":[],"x":0.8706,"y":0.8185,"text":"This survey reviews efficient test-time compute (TTC) strategies for LLMs, motivated by the observation that reasoning models apply fixed inference-time compute regardless of difficulty\u2014overthinking simple queries (e.g., '1+1=?') and underthinking hard ones. It introduces a two-tiered taxonomy distinguishing L1 controllable methods (which obey a pre-defined compute budget via a constrained optimization) from L2 adaptive methods (which dynamically scale compute by task difficulty/model confidence via a penalized objective), cross-cut by parallel vs. sequential paradigms and prompting/SFT/RL implementations. The authors empirically benchmark proprietary and open reasoning models on math/STEM/non-STEM datasets, showing RL-trained reasoning models produce up to 5\u00d7 longer outputs for SOTA accuracy while distilled compact models often produce the longest yet weakest outputs, and that thinking budgets give diminishing returns on easy tasks. They flag hybrid fast-slow thinking models and multi-modal/embodied TTC as key future directions."},{"id":"engraved.blog:b8c16edd622e","kind":"paper","n":338,"label":"Building A Virtual Machine inside ChatGPT","authors":"Degrave (Jonas Degrave)","year":2022,"url":"https://www.engraved.blog/building-a-virtual-machine-inside/","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":0,"alsoIn":[],"x":0.4051,"y":0.8875,"text":"A blog post documenting that ChatGPT (the original 2022 GPT-3.5 assistant) can simulate a fully stateful Linux virtual machine purely through a prompt, responding to shell commands with plausible filesystem, programming, Docker, and networking behavior. The author shows ChatGPT maintains state across commands (creating then retrieving files), executes code-like tasks (computing the first 10 primes), and even simulates an 'alt-internet' where it imagines an OpenAI URL hosting an 'Assistant' LLM that can itself be queried \u2014 recursively running another VM inside that imagined chatbot. The piece is an informal demonstration of emergent world-modeling/simulation in a language model rather than a controlled study."},{"id":"neurips.cc:9c7e4f09bb35","kind":"paper","n":339,"label":"TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks","authors":"Frank Xu, Yufan Song, Boxuan Li, Yuxuan Tang, Kritanjali Jain, Mengxue Bao, Zora","year":2025,"url":"https://neurips.cc/virtual/2025/poster/121705","region":"t35","hemi":"llm","lobe":"occipital","color":"#b0679e","indegree":2,"alsoIn":[],"x":0.6255,"y":0.9528,"text":"TheAgentCompany introduces an extensible benchmark for evaluating LLM agents on consequential, real-world professional tasks that a digital worker performs: browsing the web, writing code, running programs, and communicating with coworkers. The authors build a self-contained, reproducible environment with internal websites and data mimicking a small software company, populated with a variety of work tasks. They test baseline agents driven by both closed API-based and open-weights LLMs. The most competitive agent autonomously completes 30% of tasks, showing that simpler tasks are partly automatable while difficult long-horizon tasks remain beyond current systems."},{"id":"huggingface.co:e9042540da12","kind":"paper","n":340,"label":"Cosmos Predict 2.5 & Transfer 2.5: Evolving the World Foundation Models for Physical AI","authors":"NVIDIA","year":2025,"url":"https://huggingface.co/blog/nvidia/cosmos-predict-and-transfer2-5","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":7,"alsoIn":[],"x":0.4381,"y":0.9531,"text":"NVIDIA introduces Cosmos Predict 2.5 and Transfer 2.5, updates to the open Cosmos world foundation models for physical AI (robotics, autonomous vehicles, simulation). Predict 2.5 unifies the previously separate Text2World, Image2World, and Video2World models into a single architecture trained on 200 million high-quality clips, using Cosmos Reason 1 (a 7B physical-AI reasoning VLM) as its text encoder and a new RL algorithm to improve quality and prompt alignment. Transfer 2.5 performs spatially-controlled world-to-world translation at 3.5x smaller size (2B vs the prior 7B) while achieving better prompt/physics alignment and less error accumulation. Reported gains include up to 60% improvement in 3D lane and cuboid detection on generated multi-view AV videos versus Transfer1-7B-Sample-AV."},{"id":"arxiv:2503.20523","kind":"paper","n":341,"label":"GAIA-2: A Controllable Multi-View Generative World Model for Autonomous Driving","authors":"Russell et al. (Wayve)","year":2025,"url":"https://arxiv.org/abs/2503.20523","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":2,"alsoIn":[],"x":0.372,"y":0.7218,"text":"GAIA-2 is a latent diffusion generative world model for autonomous driving that produces high-resolution, spatiotemporally consistent multi-camera driving video (up to five views at 448\u00d7960) conditioned on a rich set of structured inputs\u2014ego-vehicle dynamics, 3D agent bounding boxes, scene metadata, road semantics, camera geometry\u2014plus external latent embeddings (CLIP text/image and a proprietary scenario embedding). It pairs a high-compression continuous video tokenizer (32\u00d7 spatial, 8\u00d7 temporal, 64 latent channels) with an 8.4B-parameter space-time factorized transformer trained by flow matching. The model unifies multiple inference modes (generation from scratch, autoregressive prediction, inpainting, scene editing) to simulate both common and rare safety-critical scenarios across UK, US, and German driving environments. Evaluation reports FDD, FID, FVMD, and bounding-box IoU metrics that improve over training, with validation loss found to correlate best with human perceptual preference."},{"id":"arxiv:2605.07234","kind":"paper","n":342,"label":"Reformulating KV Cache Eviction Problem for Long-Context LLM Inference","authors":"","year":null,"url":"https://arxiv.org/abs/2605.07234","region":"t18","hemi":"llm","lobe":"temporal","color":"#67b09a","indegree":2,"alsoIn":[],"x":0.8828,"y":0.4697,"text":"The paper tackles KV cache memory/latency overhead in long-context LLM inference, arguing that existing eviction methods rely only on local attention weights and ignore value representations, output projection, and inter-head interactions. It reformulates KV cache eviction as an output-aware, layer-wise matrix multiplication approximation problem and introduces LaProx, which scores tokens by the product of the L2 norms of attention columns and projected value rows (A and VW_O), enabling globally comparable, model-wide token selection instead of head-wise decisions. Across 19 datasets on LongBench and Needle-In-A-Haystack, LaProx maintains performance with only ~5% of the KV cache and consistently beats prior SOTA, achieving up to 2x accuracy-loss reduction under extreme compression with minimal overhead."},{"id":"arxiv:1910.08210","kind":"paper","n":343,"label":"RTFM: Generalising to Novel Environment Dynamics via Reading","authors":"Zhong, Rockt\u00e4schel & Grefenstette","year":2019,"url":"https://arxiv.org/abs/1910.08210","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":2,"alsoIn":[],"x":0.1557,"y":0.4443,"text":"The paper introduces RTFM (Read to Fight Monsters), a grounded RL problem where an agent must jointly reason over a language goal, a document describing procedurally-generated environment dynamics, and grid-world observations, requiring up to five hops of cross-referencing reasoning to win. To prevent memorization, environment dynamics (monster-team-modifier-element assignments) and natural-language templated descriptions are procedurally generated, yielding millions of distinct dynamics that differ between train and eval. The authors propose txt2\u03c0, built on a Bidirectional FiLM (FiLM\u00b2) layer that captures three-way interactions by modulating visual features with text and text features with vision. txt2\u03c0 generalizes to unseen dynamics, outperforms FiLM and language-conditioned CNN baselines in sample efficiency and win rate, and via curriculum learning solves the full RTFM, though it still trails human players."},{"id":"arxiv:2309.12288","kind":"paper","n":344,"label":"The Reversal Curse: LLMs Trained on 'A is B' Fail to Learn 'B is A'","authors":"Berglund, Tong, Kaufmann, Balesni, Stickland, Korbak & Evans","year":2024,"url":"https://arxiv.org/abs/2309.12288","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":0,"alsoIn":[],"x":0.7662,"y":0.7141,"text":"The paper exposes the 'Reversal Curse': auto-regressive LLMs trained on a fact of the form 'A is B' fail to generalize to the logically equivalent 'B is A'. Through controlled finetuning experiments on fictitious facts with GPT-3 and Llama-1, models answer correctly when the test order matches training but score near 0% (and assign no higher log-probability to the correct name than a random one) when the order is reversed; this holds across model sizes, families, paraphrase augmentation, larger datasets, and prompt tuning. A complementary zero-shot test on 1,573 real celebrity parent-child pairs shows GPT-4 identifies a celebrity's parent 79% of the time but the parent's child only 33% of the time. The curse does not apply to in-context learning, where models reverse facts at ~100% accuracy."},{"id":"arxiv:2412.02252","kind":"paper","n":345,"label":"Compressing KV Cache for Long-Context LLM Inference with Inter-Layer Attention Similarity","authors":"","year":null,"url":"https://arxiv.org/abs/2412.02252","region":"t18","hemi":"llm","lobe":"temporal","color":"#67b09a","indegree":2,"alsoIn":[],"x":0.8724,"y":0.4535,"text":"Long-context LLM inference is bottlenecked by KV cache memory that grows linearly with context length, and existing compression methods discard tokens, losing critical information. The paper proposes POD (Proximal tokens over Distant tokens), which keeps the full KV cache for proximal (initial + recent) tokens but shares key states across consecutive similar layers for distant tokens \u2014 exploiting the observation that attention scores for distant tokens are highly redundant between adjacent layers \u2014 instead of dropping them. An offline JS-divergence analysis groups layers head-wise into sharing blocks, followed by a lightweight post-training adaptation (5B Dolma tokens). POD cuts KV cache memory by up to 35% with essentially no performance loss, beats token-eviction baselines, and is orthogonal to token-selection methods like SnapKV."},{"id":"arxiv:2407.01392","kind":"paper","n":346,"label":"Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion","authors":"Chen, Mart\u00ed Monso, Du, Simchowitz, Tedrake & Sitzmann","year":2024,"url":"https://arxiv.org/abs/2407.01392","region":"t32","hemi":"wm","lobe":"frontal","color":"#8b4f9a","indegree":2,"alsoIn":[],"x":0.3274,"y":0.1517,"text":"The paper introduces Diffusion Forcing (DF), a training paradigm where a diffusion model denoises a set of tokens each carrying an independent, per-token noise level, reframing noise as a form of partial masking. Instantiated as Causal Diffusion Forcing (CDF) with a causal RNN, it trains a next-token prediction model that combines variable-length generation (from next-token models) with guidance toward desirable trajectories (from full-sequence diffusion), and samples via a 2D noise-scheduling matrix. Empirically it stabilizes long-horizon autoregressive video rollouts far past the training horizon where baselines diverge, and\u2014via a new Monte Carlo Guidance scheme\u2014outperforms Diffuser and offline-RL baselines on D4RL maze planning and enables memory-based, robust visuomotor imitation learning. The authors also prove the objective optimizes a (reweighted) variational lower bound on the likelihoods of all subsequences/noise-level sequences."},{"id":"arxiv:2503.20552","kind":"paper","n":347,"label":"Injecting Adrenaline into LLM Serving: Boosting Resource Utilization and Throughput via Attention Disaggregation","authors":"","year":null,"url":"https://arxiv.org/abs/2503.20552","region":"t24","hemi":"llm","lobe":"frontal","color":"#6784b0","indegree":4,"alsoIn":[],"x":0.9277,"y":0.5767,"text":"LLM serving with prefill-decoding (PD) disaggregation wastes GPU resources: prefill instances have low memory (HBM) utilization while decoding instances have low compute utilization. The paper proposes Adrenaline, which disaggregates part of the memory-bound decoding-phase attention computation and offloads it to a remote attention executor colocated on prefill instances, raising prefill memory utilization and enabling larger decoding batch sizes for higher compute utilization. It uses three techniques: low-latency decoding synchronization (CUDA-graph kernel pre-launching), resource-efficient prefill colocation (SM partitioning via NVIDIA MPS), and load-aware offloading scheduling that bounds the offloading ratio. Built on vLLM, it achieves up to 2.28x HBM capacity, 2.07x HBM bandwidth utilization in prefill instances, 1.67x decoding compute utilization, and 1.68x higher throughput."},{"id":"arxiv:1706.02275","kind":"paper","n":348,"label":"Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments (MADDPG)","authors":"Lowe et al.","year":2017,"url":"https://arxiv.org/abs/1706.02275","region":"t22","hemi":"wm","lobe":"temporal","color":"#4f849a","indegree":3,"alsoIn":[],"x":0.1356,"y":0.2898,"text":"The paper addresses why traditional RL fails in multi-agent settings: Q-learning suffers from environment non-stationarity (breaking experience replay) and policy gradients have variance that grows exponentially with agent count. It proposes MADDPG, an actor-critic method using centralized training with decentralized execution, where each agent's critic is augmented with all agents' observations and actions while actors use only local observations at test time. MADDPG reliably learns coordination on cooperative and competitive particle-world tasks where DQN, Actor-Critic, TRPO, DDPG, and REINFORCE fail, e.g. reaching the target 84.0% of episodes in cooperative communication vs 32.0% for DDPG."},{"id":"philpapers.org:4f7ed9918446","kind":"paper","n":349,"label":"Behavior, Purpose and Teleology","authors":"Arturo Rosenblueth, Norbert Wiener, Julian Bigelow","year":1943,"url":"https://philpapers.org/rec/ROSBPA","region":"t2","hemi":"llm","lobe":"temporal","color":"#b07d67","indegree":2,"alsoIn":[],"x":0.6166,"y":0.7333,"text":"The provided text is not a paper but a Cloudflare security block page preventing access to philpapers.org. No paper content, methods, results, or findings are available for indexing."},{"id":"arxiv:2307.13692","kind":"paper","n":350,"label":"ARB: Advanced Reasoning Benchmark for Large Language Models","authors":"Tomohiro Sawada et al.","year":2023,"url":"https://arxiv.org/abs/2307.13692","region":"t33","hemi":"llm","lobe":"parietal","color":"#ad67b0","indegree":3,"alsoIn":[],"x":0.811,"y":0.6071,"text":"Existing LLM reasoning benchmarks are saturating before models reach expert performance, so the paper introduces ARB (Advanced Reasoning Benchmark), a dataset of graduate/professional-level problems in mathematics, physics, chemistry, biology, and law drawn from sources unlikely to be in training data. It evaluates GPT-4, ChatGPT (gpt-3.5-turbo), text-davinci-003, and Claude using chain-of-thought prompting, showing current models score well below 50% on the quantitative (numerical/symbolic/proof-like) subsets. The authors also propose a rubric-based self-evaluation method where GPT-4 generates a 10-point rubric from a reference solution and grades intermediate reasoning steps, finding moderately-high agreement (Pearson 0.78\u20130.91) with human annotators."},{"id":"arxiv:2308.09723","kind":"paper","n":351,"label":"FineQuant: Unlocking Efficiency with Fine-Grained Weight-Only Quantization for LLMs","authors":"","year":null,"url":"https://arxiv.org/abs/2308.09723","region":"t37","hemi":"llm","lobe":"parietal","color":"#b06788","indegree":4,"alsoIn":[],"x":0.8873,"y":0.3627,"text":"The paper tackles the memory-bandwidth bottleneck and high deployment cost of large language model inference by proposing FineQuant, a weight-only, post-training quantization method that needs no fine-tuning or calibration data. Its core is an adaptive fine-grained heuristic that uses only the pre-trained weights to decide per-matrix quantization granularity (recursively halving group size when a group's value range grows beyond a threshold \u03b1), paired with efficient CUTLASS GPU GEMM kernels that fuse dequantization with fp16/bf16\u00d7int8/int4 matrix multiplication. Applied to dense models (up to OPT-175B) and internal MoE models, it preserves accuracy down to 4-bit (and 3-bit for some parts) while cutting model size. It lets OPT-175B run on just 2 GPUs (64% cost reduction) and delivers up to 3.65\u00d7 higher throughput on the same number of GPUs."},{"id":"arxiv:2509.00579","kind":"paper","n":352,"label":"KVComp: A High-Performance, LLM-Aware, Lossy Compression Framework for KV Cache","authors":"","year":null,"url":"https://arxiv.org/abs/2509.00579","region":"t18","hemi":"llm","lobe":"temporal","color":"#67b09a","indegree":3,"alsoIn":[],"x":0.9098,"y":0.4072,"text":"Long-context LLM inference is bottlenecked by the KV cache, which can exceed the model's own weight memory (e.g., LLaMA2-30B at 32K context, batch 8 produces >100GB of KV cache vs 60GB weights). KVComp is an LLM-aware lossy compression framework that combines fine-grained error-controlled quantization (block-wise/channel-wise for K, token-wise for V) with GPU-optimized Huffman entropy encoding and a branch-divergence-free, cache-resident decompression that fuses decoding directly with the attention matrix-vector multiplication. It achieves on average 47% and up to 83% higher memory reduction than KIVI with little/no accuracy loss, sustains >400 GB/s (K) and >180 GB/s (V) fused throughput, and at long context lengths even outperforms cuBLAS attention kernels due to reduced data movement."},{"id":"arxiv:2310.03744","kind":"paper","n":353,"label":"Improved Baselines with Visual Instruction Tuning","authors":"Haotian Liu, Chunyuan Li, Yuheng Li, Yong Jae Lee","year":2023,"url":"https://arxiv.org/abs/2310.03744","region":"t30","hemi":"llm","lobe":"temporal","color":"#8c67b0","indegree":3,"alsoIn":[],"x":0.8167,"y":0.2078,"text":"This paper presents a systematic study of design choices for large multimodal models (LMMs) under the LLaVA framework, showing the simple fully-connected vision-language connector is powerful and data-efficient. Two key modifications\u2014swapping to a two-layer MLP connector with CLIP-ViT-L-336px and adding academic-task-oriented VQA data with response-formatting prompts\u2014yield LLaVA-1.5, which reaches SOTA across 11-12 benchmarks. The 13B model uses only 1.2M public data samples (558K pretrain + 665K finetune) and trains in ~1 day on a single 8-A100 node. The paper also explores open problems: scaling to higher resolution via image gridding (LLaVA-1.5-HD), data efficiency, hallucination reduction, and compositional/multilingual generalization."},{"id":"arxiv:1812.00922","kind":"paper","n":354,"label":"Multi-agent Deep Reinforcement Learning with Extremely Noisy Observations","authors":"","year":null,"url":"https://arxiv.org/abs/1812.00922","region":"t22","hemi":"wm","lobe":"temporal","color":"#4f849a","indegree":4,"alsoIn":[],"x":0.1317,"y":0.2992,"text":"The paper tackles cooperative multi-agent reinforcement learning under partial observability where most agents' observations are extremely noisy and only weakly correlated with the true state, making standard CTDE methods fail. It proposes MADDPG-M, which augments MADDPG with a learned communication medium and a two-level, two-time-scale hierarchy: top-level communication policies (run every C steps, trained on accumulated extrinsic rewards) decide whose observations to share, while bottom-level action policies are trained decentrally using intrinsic rewards defined relative to the shared medium. Across six progressively harder Cooperative Navigation variants (broadcasting and unicasting; fixed/alternating/dynamic), MADDPG-M nearly matches the hard-coded optimal-communication upper bound (DDPG-OC) and far exceeds DDPG, MADDPG, and a Meta-agent. The agents learn what and when to communicate purely from indirect environmental feedback, with no explicit reward for communication actions."},{"id":"arxiv:2305.18654","kind":"paper","n":355,"label":"Faith and Fate: Limits of Transformers on Compositionality","authors":"Dziri, Lu, Sclar, Li et al.","year":2023,"url":"https://arxiv.org/abs/2305.18654","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":0,"alsoIn":[],"x":0.8066,"y":0.7775,"text":"The paper investigates fundamental limits of autoregressive transformer LLMs (GPT3, ChatGPT, GPT4) on three compositional tasks\u2014multi-digit multiplication, Einstein's logic grid puzzles, and a dynamic programming problem\u2014by formulating each task as a computation graph to quantify complexity (depth, width, parallelism). Empirically, models achieve near-perfect accuracy on low-complexity/in-domain instances but collapse toward zero as problem size grows, even with scratchpad reasoning, exhaustive finetuning, or grokking-style extended training. The authors argue transformers solve such tasks via 'linearized subgraph matching' (pattern matching against memorized sub-computations) rather than learning generalizable algorithms, and provide theoretical proofs that error probability grows exponentially with the number of composed reasoning steps. The conclusion is that out-of-the-box transformers are inherently limited on compositionally complex, multi-hop tasks."},{"id":"arxiv:2301.08243","kind":"paper","n":356,"label":"Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture","authors":"Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Mi","year":2023,"url":"https://arxiv.org/abs/2301.08243","region":"t23","hemi":"wm","lobe":"occipital","color":"#4f799a","indegree":6,"alsoIn":[],"x":0.2211,"y":0.4021,"text":"The paper introduces I-JEPA (Image-based Joint-Embedding Predictive Architecture), a non-generative self-supervised learning method that learns semantic image representations without hand-crafted data augmentations. From a single context block, it predicts the representations (not pixels) of multiple target blocks in the same image, using a ViT context-encoder, a narrow ViT predictor, and an EMA-updated target-encoder, with a multi-block masking strategy as the key design choice. I-JEPA outperforms pixel-reconstruction methods like MAE on ImageNet linear probing, low-shot, and transfer tasks while being far more compute-efficient, and matches view-invariance methods on semantic tasks while beating them on low-level tasks like object counting and depth prediction."},{"id":"philpapers.org:8c0576aa7f12","kind":"paper","n":357,"label":"The Predictive Mind","authors":"Jakob Hohwy","year":2013,"url":"https://philpapers.org/rec/HOHTPM-2","region":"t2","hemi":"llm","lobe":"temporal","color":"#b07d67","indegree":6,"alsoIn":[],"x":0.6136,"y":0.7313,"text":"The provided paper text could not be retrieved: the supplied content is a Cloudflare access-denied/block page from philpapers.org, not the actual document. Based solely on the title, 'The Predictive Mind' is the well-known work by Jakob Hohwy articulating the predictive processing / prediction-error-minimization (Bayesian brain) account of perception, action and cognition. However, no body text, methods, experiments, or numerical results were available in the input, so no grounded claims or metrics can be extracted. This record is a placeholder reflecting that the source content was blocked."},{"id":"arxiv:2406.03689","kind":"paper","n":358,"label":"Evaluating the World Model Implicit in a Generative Model","authors":"Vafa, Chen, Rambachan, Kleinberg & Mullainathan","year":2024,"url":"https://arxiv.org/abs/2406.03689","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":1,"alsoIn":[],"x":0.2716,"y":0.6004,"text":"The paper addresses how to evaluate whether generative sequence models implicitly recover a coherent world model, formalizing the underlying reality as a deterministic finite automaton (DFA). Inspired by the Myhill-Nerode theorem, it introduces two model-agnostic evaluation metrics \u2014 sequence compression (do sequences reaching the same state admit the same continuations?) and sequence distinction (do sequences reaching different states have distinct continuations?) \u2014 that probe the implied Myhill-Nerode boundary rather than just next-token validity. Applied to transformers trained on NYC taxi navigation, Othello, and LLMs solving logic puzzles, the metrics reveal that models scoring near-perfectly on existing diagnostics (next-token validity, state probes) have far less coherent world models than they appear. This incoherence creates fragility: e.g., navigation models that find shortest paths break down when detours are introduced."},{"id":"arxiv:2405.03520","kind":"paper","n":359,"label":"Is Sora a World Simulator? A Comprehensive Survey on General World Models and Beyond","authors":"Zheng Zhu, Xiaofeng Wang et al. (GigaAI)","year":2024,"url":"https://arxiv.org/abs/2405.03520","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":2,"alsoIn":[],"x":0.4414,"y":0.7923,"text":"This survey comprehensively reviews general world models\u2014models that understand the world by predicting the future\u2014using OpenAI's Sora as a focal point and asking whether it qualifies as a true world simulator. It organizes the field into three application directions: video generation (GAN, diffusion, autoregressive, and masked-modeling techniques, plus Sora's likely DiT-based architecture), autonomous-driving world models (end-to-end driving and 2D/3D neural driving simulators), and autonomous-agent world models (RSSM/JEPA/Transformer/diffusion structures for games and robotics). It catalogs representative methods, datasets, and benchmarks with comparative result tables, then argues video generation alone is not equivalent to world modeling and lays out open challenges (causal/counterfactual reasoning, physical-law compliance, generalization, computational efficiency, evaluation) and ethical concerns."},{"id":"introl.com:c695c416f507","kind":"paper","n":360,"label":"World Models Race 2026","authors":"Introl","year":2026,"url":"https://introl.com/blog/world-models-race-agi-2026","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":1,"alsoIn":[],"x":0.3274,"y":0.9072,"text":"This is an industry landscape article surveying the surge of the 'world models' paradigm in late 2025\u2013early 2026, framed as an alternative to large language models for achieving general intelligence. It profiles major players and products \u2014 Yann LeCun's AMI Labs (raising \u20ac500M at a \u20ac3B pre-launch valuation), DeepMind's Genie 3 (real-time interactive 3D world model), World Labs' Marble (first commercial world-model product), and NVIDIA's Cosmos platform \u2014 alongside the convergence of video generators (Sora 2, Runway Gen-4.5) toward physics-aware world simulation. The piece argues LLMs are fundamentally limited by lacking grounding in physical reality, and details the heavier compute/storage infrastructure (8\u201332 GPUs per inference, petabyte-scale video storage) world models demand. It targets infrastructure planners weighing the shift from text processing to video generation, physics simulation, and embodied reasoning."},{"id":"arxiv:2603.13647","kind":"paper","n":361,"label":"PLUME: Building a Network-Native Foundation Model for Wireless Traces via Protocol-Aware Tokenization","authors":"","year":null,"url":"https://arxiv.org/abs/2603.13647","region":"t39","hemi":"llm","lobe":"occipital","color":"#9a4f5b","indegree":1,"alsoIn":[],"x":0.8395,"y":0.5654,"text":"Wireless packet analysis suffers because general LLMs and standard tokenizers (BPE, byte-level) flatten 802.11 traces into strings, destroying field boundaries, timing, and protocol hierarchy. The authors present Plume, a compact 140M-parameter decoder-only (GPT-2 backbone) foundation model trained from scratch on curated Wireshark PDML dissections using a protocol- and timing-aware tokenizer plus an HDBSCAN+MMR curation pipeline. On five real-world 802.11 failure categories Plume reaches 74.1\u201397.3% next-packet token accuracy, AUROC\u22650.99 for zero-shot anomaly detection, and 73.2% five-class root-cause accuracy, matching or exceeding Claude Opus 4.6 and GPT-5.4 with >600\u00d7 fewer parameters at effectively zero marginal cost. The work argues that representation/tokenization and data quality, not scale, are the decisive levers for network-native modeling."},{"id":"arxiv:2410.05229","kind":"paper","n":362,"label":"GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models","authors":"Mirzadeh, Alizadeh, Shahrokhi, Tuzel, Bengio & Farajtabar","year":2024,"url":"https://arxiv.org/abs/2410.05229","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":0,"alsoIn":[],"x":0.8589,"y":0.7503,"text":"The paper investigates whether reported LLM gains on the GSM8K grade-school math benchmark reflect genuine reasoning, and introduces GSM-Symbolic, a benchmark that generates many variants of each question from symbolic templates (100 templates \u00d7 50 samples = 5000 examples per set). Evaluating 25 open and closed models, the authors find LLMs show large accuracy variance across instantiations of the same question and consistently drop in accuracy when only numbers change or when extra clauses are added. They further introduce GSM-NoOp, which inserts seemingly-relevant but inconsequential clauses, causing catastrophic accuracy drops (up to ~65%) across all models. They conclude current LLM math 'reasoning' is fragile, in-distribution pattern matching rather than formal logical reasoning."},{"id":"arxiv:2405.05925","kind":"paper","n":363,"label":"FuXi-ENS: A machine learning model for medium-range ensemble weather forecasting","authors":"Xiaohui Zhong, Lei Chen, Hao Li, et al.","year":2024,"url":"https://arxiv.org/abs/2405.05925","region":"t34","hemi":"wm","lobe":"temporal","color":"#9a4f93","indegree":3,"alsoIn":[],"x":0.2444,"y":0.1323,"text":"Applying ML to ensemble weather forecasting is hard because prior ML approaches (GenCast, SEEDS) depend on EDA or operational NWP ensemble members and run at coarse resolution. The authors introduce FuXi-ENS, a VAE-based autoregressive model that generates 6-hourly global ensemble forecasts up to 15 days at 0.25\u00b0 resolution (5 upper-air variables at 13 pressure levels plus 13 surface variables), using only ERA5 analysis for initialization plus a learned perturbation module that injects flow-dependent perturbations at initial conditions and every forecast step. Its key innovation is a loss combining CRPS with KL divergence (replacing the standard L1+KL VAE loss). FuXi-ENS beats the 51-member ECMWF ensemble in CRPS for 98.1% of 360 variable/lead-time combinations and generates a 15-day member in ~10 seconds on an A100 GPU."},{"id":"arxiv:2405.05967","kind":"paper","n":364,"label":"Diffusion Models with f-Divergence Distribution Matching (Distilling Diffusion Models into Conditional GANs)","authors":"Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Pa","year":2024,"url":"https://arxiv.org/abs/2405.05967","region":"t32","hemi":"wm","lobe":"frontal","color":"#8b4f9a","indegree":3,"alsoIn":[],"x":0.3984,"y":0.1065,"text":"The paper tackles the slow multi-step inference of text-to-image diffusion models by distilling a pre-trained diffusion model into a single-step conditional GAN (Diffusion2GAN). It reframes distillation as paired noise-to-image translation using noise-image pairs collected from the teacher's deterministic ODE trajectory, trained with a novel latent-space perceptual loss (E-LatentLPIPS) plus a multi-scale conditional diffusion discriminator. The one-step generator synthesizes a 512px image in 0.09s (vs 2.59s for the 50-step SD1.5 teacher) and outperforms prior one-step distillation methods, e.g. FID-30k 9.29 on COCO2014 vs 11.49 for DMD and 13.10 for InstaFlow-0.9B. When distilling SDXL it beats SDXL-Turbo and SDXL-Lightning on FID and CLIP-score (FID-5k 25.49, CLIP 0.347 on COCO2017)."},{"id":"arxiv:2411.14499","kind":"paper","n":365,"label":"Understanding World or Predicting Future? A Comprehensive Survey of World Models","authors":"Ding et al.","year":2024,"url":"https://arxiv.org/abs/2411.14499","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":11,"alsoIn":[],"x":0.3533,"y":0.8682,"text":"This survey reviews the literature on world models, framing them around two primary functions: constructing internal/implicit representations to understand the world's mechanisms (e.g., model-based RL, JEPA, world knowledge in LLMs) and predicting future states to simulate and guide decision-making (e.g., video generation, embodied environments). It proposes a systematic categorization spanning Section 3 (implicit representation) and Section 4 (future prediction), then maps how four application domains \u2014 generative games, embodied/robotics, autonomous driving/urban, and social simulacra \u2014 emphasize different aspects. It concludes that purely data-driven generative scaling is insufficient to recover robust physical laws and outlines open problems including counterfactual/physics fidelity, social-dimension modeling, benchmarking, sim-to-real, efficiency, and ethics/safety."},{"id":"arxiv:2304.09116","kind":"paper","n":366,"label":"NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers","authors":"Kai Shen et al.","year":2023,"url":"https://arxiv.org/abs/2304.09116","region":"t11","hemi":"llm","lobe":"occipital","color":"#81b067","indegree":0,"alsoIn":[],"x":0.7288,"y":0.2583,"text":"Large-scale TTS systems that quantize speech into discrete tokens and model them autoregressively suffer from unstable prosody, word skipping/repeating, and poor voice quality. NaturalSpeech 2 instead uses a neural audio codec with residual vector quantizers to produce continuous latent vectors and a non-autoregressive latent diffusion model (WaveNet-based score network) to generate them conditioned on a phoneme encoder plus duration/pitch predictors, with a speech-prompting mechanism for in-context learning. Scaled to ~400M parameters and 44K hours of speech/singing data, it outperforms prior zero-shot systems (YourTTS, VALL-E) in prosody/timbre similarity, robustness, and naturalness on unseen speakers. It also achieves novel zero-shot singing synthesis from only a speech prompt, and extends to voice conversion and speech enhancement."},{"id":"arxiv:2403.04253","kind":"paper","n":367,"label":"Mastering Memory Tasks with World Models (Recall to Imagine, R2I)","authors":"Samsami et al.","year":2024,"url":"https://arxiv.org/abs/2403.04253","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":0,"alsoIn":[],"x":0.1874,"y":0.3146,"text":"Current model-based RL (MBRL) agents rely on RNN or Transformer world-model backbones that struggle with long-term memory and long-horizon credit assignment. The paper introduces Recall to Imagine (R2I), the first MBRL method to integrate a parallelizable variant of structured state space models (S4/S5-style SSMs) into DreamerV3's world model, forming a Structured State-Space Model (S3M) that trains via parallel scan and runs recurrently at inference. R2I sets new state-of-the-art on memory and credit-assignment benchmarks (BSuite, POPGym) and achieves superhuman performance in the 3D Memory Maze, while matching DreamerV3 on standard Atari and DMC tasks. It also runs up to 9x faster than DreamerV3, yielding faster wall-clock convergence."},{"id":"arxiv:2606.18250","kind":"paper","n":368,"label":"Future Dynamic 3D Reconstruction: A 3D World Model with Disentangled Ego-Motion","authors":"Future Dynamic 3D Reconstruction authors","year":2026,"url":"https://arxiv.org/abs/2606.18250","region":"t15","hemi":"wm","lobe":"occipital","color":"#4f9a62","indegree":4,"alsoIn":[],"x":0.3643,"y":0.707,"text":"Generative 2D video world models achieve photorealism but mix ego-motion with scene dynamics in the image plane, causing physical inconsistencies (morphing/vanishing objects) over long horizons. This paper introduces FR3D, a 3D world model that predicts a persistent 3D latent representation for future dynamic 3D reconstruction from monocular images, explicitly disentangling inferred ego-motion (treated as a latent action proxy) from world-motion via two cross-attention-coupled masked transformers (pose + spatial) operating in CUT3R's latent space. Trained only on Waymo via teacher-student distillation from the frozen CUT3R foundation model, FR3D generalizes zero-shot to KITTI and nuScenes, beating DINO-Foresight and CUT3R-Foresight baselines on depth and pose forecasting even 2 seconds into the future. It also runs far cheaper than 2D video model Vista (5.75 GB vs 11.60 GB VRAM, 0.85s vs 24.57s latency)."},{"id":"pubmed.ncbi.nlm.nih.gov:9eeb70dd85f3","kind":"paper","n":369,"label":"Model-based influences on humans' choices and striatal prediction errors","authors":"Daw, N. D., Gershman, S. J., Seymour, B., Dayan, P. & Dolan, R. J.","year":2011,"url":"https://pubmed.ncbi.nlm.nih.gov/21435563/","region":"t5","hemi":"llm","lobe":"parietal","color":"#b09e67","indegree":0,"alsoIn":[],"x":0.5568,"y":0.6772,"text":"The paper investigates whether the human mesostriatal dopamine system is a pure model-free reinforcement learner or also reflects model-based (goal-directed/planning) computations. The authors designed a two-stage (multistep) Markov decision task with fMRI, in which model-based and model-free influences on choice could be dissociated via a reward \u00d7 transition-probability factorial analysis, and tested whether the ventral striatal BOLD prediction-error signal was purely model-free. Choices reflected both model-free and model-based influences, and contrary to expectations the ventral striatal BOLD reward-prediction-error (RPE) signal reflected BOTH types of prediction in proportions matching those that best explained each subject's behavior. The findings challenge the idea of a separate model-free learner and argue for a more integrated computational architecture underlying human decision-making."},{"id":"arxiv:2303.11366","kind":"paper","n":370,"label":"Reflexion: Language Agents with Verbal Reinforcement Learning","authors":"Shinn et al.","year":2023,"url":"https://arxiv.org/abs/2303.11366","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":1,"alsoIn":[],"x":0.7108,"y":0.827,"text":"Language agents built on LLMs struggle to learn from trial-and-error because traditional RL needs many samples and costly fine-tuning. Reflexion reinforces agents through verbal feedback instead of weight updates: agents convert binary/scalar task feedback into self-reflective natural-language summaries stored in an episodic memory buffer that conditions decisions in subsequent trials. Using Actor, Evaluator, and Self-Reflection LLM modules, it yields large gains across decision-making (AlfWorld +22% absolute), reasoning (HotPotQA +20%), and coding tasks. Notably it reaches 91% pass@1 on HumanEval Python, beating GPT-4's 80%."},{"id":"arxiv:2503.16416","kind":"paper","n":371,"label":"A Survey on Evaluation of LLM-based Agents","authors":"Yehudai et al.","year":2025,"url":"https://arxiv.org/abs/2503.16416","region":"t35","hemi":"llm","lobe":"occipital","color":"#b0679e","indegree":8,"alsoIn":[],"x":0.6086,"y":0.9523,"text":"This paper presents the first comprehensive survey of evaluation methods for LLM-based agents, which extend static LLMs into autonomous systems that plan, reason, use tools, and act in dynamic environments. It organizes the field across five perspectives: core agent capabilities (planning, tool use, self-reflection, memory), application-specific benchmarks (web, software-engineering, scientific, and conversational agents), generalist-agent evaluation, a cross-cutting analysis of benchmark dimensions (data curation, environment, interface, metric, safety), and developer-facing evaluation frameworks. It documents a clear trend toward more realistic, dynamic, continuously updated ('live') benchmarks and identifies critical gaps in assessing cost-efficiency, safety/robustness, and fine-grained, scalable evaluation. The authors also recommend de facto benchmarks per domain and call for decoupling LLM-backbone evaluation from agent-harness evaluation."},{"id":"arxiv:2108.13343","kind":"paper","n":372,"label":"A Mathematical Walkthrough and Discussion of the Free Energy Principle","authors":"","year":null,"url":"https://arxiv.org/abs/2108.13343","region":"t25","hemi":"llm","lobe":"parietal","color":"#6779b0","indegree":7,"alsoIn":[],"x":0.6031,"y":0.8181,"text":"This paper provides a mathematically detailed yet accessible walkthrough of the Free Energy Principle (FEP), a theory connecting self-organization in stochastic dynamical systems to variational Bayesian inference. The authors derive the core results step-by-step: from Langevin dynamics with a non-equilibrium steady state (NESS), through the Helmholtz/Ao decomposition and Markov blanket partitioning, to the identification of internal state dynamics with variational free energy minimization under the Laplace approximation. They also discuss the Expected Free Energy for active inference and provide an extensive appendix critically examining the many restrictive assumptions underlying the FEP and ongoing controversies in the community."},{"id":"arxiv:2212.06817","kind":"paper","n":373,"label":"RT-1: Robotics Transformer for Real-World Control at Scale","authors":"Brohan, Brown, Carbajal et al.","year":2022,"url":"https://arxiv.org/abs/2212.06817","region":"t0","hemi":"wm","lobe":"frontal","color":"#9a4f4f","indegree":2,"alsoIn":[],"x":0.1848,"y":0.681,"text":"RT-1 addresses whether a single high-capacity model can absorb large, diverse multi-task robot data and generalize zero-shot to new tasks, objects, and environments, while still running fast enough for real-time control. The method tokenizes a history of 6 images (via a FiLM-conditioned ImageNet-pretrained EfficientNet-B3) and a USE language embedding, compresses tokens with TokenLearner, and feeds a decoder-only Transformer that outputs discretized arm/base action tokens, trained via behavioral cloning on ~130k human demonstrations (744 tasks, 13 robots, 17 months). RT-1 performs over 700 instructions at 97% success and beats prior architectures on generalization to unseen tasks, distractors, and backgrounds, while running at 3 Hz. It also absorbs heterogeneous data (simulation and another robot's data) without hurting original-task performance and supports 50-step SayCan long-horizon plans."},{"id":"arxiv:2505.24618","kind":"paper","n":374,"label":"Distributed Intelligence in the Computing Continuum with Active Inference","authors":"Victor Casamayor Pujol, Boris Sedlak, Tommaso Salvatori, Karl Friston, Schahram ","year":2025,"url":"https://arxiv.org/abs/2505.24618","region":"t39","hemi":"wm","lobe":"occipital","color":"#9a4f5b","indegree":1,"alsoIn":[],"x":0.1149,"y":0.3243,"text":"The paper addresses the challenge of distributed, autonomous service management in the Computing Continuum (CC) by deploying Active Inference (AIF) agents\u2014each modeled as a POMDP\u2014to supervise individual services in a video stream processing pipeline. It introduces SLOiDs (Service Level Objectives in Devices), device-aware non-functional requirements that account for hardware heterogeneity. Experiments show AIF agents achieve over 90% SLOiD fulfillment with expert-defined transition models and around 80% when learning models online, performing comparably to extensively trained MARL agents while requiring far less data."},{"id":"arxiv:1510.00056","kind":"paper","n":375,"label":"Single-Seed Cascades on Clustered Networks","authors":"","year":null,"url":"https://arxiv.org/abs/1510.00056","region":"t36","hemi":"llm","lobe":"frontal","color":"#b06793","indegree":3,"alsoIn":[],"x":0.6271,"y":0.7452,"text":"The paper analyzes Watts' threshold cascade process started from a single randomly chosen active seed node on clustered random networks (Newman/Miller stubs-and-corners networks where every edge belongs to a triangle). Using a 'random cactus' layout, the author formulates a two-phase, two-type branching process over node-pairs (linked sibling pairs) to derive a fixed-point equation implicitly giving the extinction (non-global-cascade) probability. The analysis recovers, as a special case, the global-cascade condition of Hackett et al. and, via simulation on 10,000-node networks, shows clustering lowers cascade probability at low thresholds (smaller giant component) but raises it at higher thresholds by enabling 2-vulnerable nodes to spread the cascade."},{"id":"arxiv:2506.01799","kind":"paper","n":376,"label":"WorldExplorer: Towards Generating Fully Navigable 3D Scenes","authors":"Manuel-Andreas Schneider, Lukas H\u00f6llein, Matthias Nie\u00dfner","year":2025,"url":"https://arxiv.org/abs/2506.01799","region":"t15","hemi":"llm","lobe":"occipital","color":"#4f9a62","indegree":2,"alsoIn":[],"x":0.5777,"y":0.1291,"text":"WorldExplorer addresses text-to-3D scene generation where prior methods produce stretched/noisy artifacts when the camera moves beyond central or panoramic viewpoints. The method first builds a 360\u00b0 panoramic scene scaffold (4 Flux-generated images plus 4 inpainted views via monocular depth unprojection), then autoregressively expands it by running a camera-guided video diffusion model (Stable Virtual Camera/SEVA) along 4 pre-defined short trajectories from each of 8 start images (32 videos total), using a novel scene-memory conditioning mechanism and a depth-based collision detector. All generated images are fused into a 3D Gaussian Splatting scene initialized from a VGGT point cloud. It achieves the highest CLIP score (25.94) and the best user-study ratings (PQ 4.04, 3DC 4.02) versus six baselines, enabling fully navigable real-time rendering."},{"id":"arxiv:2412.03572","kind":"paper","n":377,"label":"Navigation World Models","authors":"Amir Bar, Gaoyue Zhou, Danny Tran, Trevor Darrell, Yann LeCun","year":2024,"url":"https://arxiv.org/abs/2412.03572","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":1,"alsoIn":[],"x":0.2989,"y":0.8187,"text":"The paper introduces the Navigation World Model (NWM), a controllable video generation model that predicts future egocentric visual observations from past observations and navigation actions (translation + yaw + time-shift), and uses it for planning via Model Predictive Control with the Cross-Entropy Method. NWM is built on a novel Conditional Diffusion Transformer (CDiT) scaled to 1B parameters and trained on diverse robot navigation datasets (SCAND, TartanDrive, RECON, HuRoN) plus unlabeled Ego4D video. It plans goal-conditioned trajectories from scratch or ranks an external policy's (NoMaD) trajectories, achieving state-of-the-art standalone navigation, and uses learned visual priors to imagine trajectories in unfamiliar environments from a single image. Training on unlabeled Ego4D improves prediction in unseen environments."},{"id":"arxiv:2501.16411","kind":"paper","n":378,"label":"PhysBench: Benchmarking and Enhancing Vision-Language Models for Physical World Understanding","authors":"Chow, Mei, Ren et al.","year":2025,"url":"https://arxiv.org/abs/2501.16411","region":"t12","hemi":"llm","lobe":"frontal","color":"#5e9a4f","indegree":4,"alsoIn":[],"x":0.6959,"y":0.2909,"text":"VLMs show strong common-sense reasoning but poor physical-world understanding, a gap critical for embodied AI. The paper introduces PhysBench, a benchmark of 10,002 interleaved video-image-text multiple-choice entries spanning 4 domains (object property, object relationships, scene understanding, physics-based dynamics), 19 sub-classes and 8 capability dimensions, evaluated on 75 VLMs. It finds models max out near human-far performance (GPT-4o best at 49.49% vs human 95.87%), with failures driven by perception errors and missing physical knowledge rather than model/data/frame scaling. It proposes PhysAgent, a framework combining VLMs with vision foundation models plus a physics knowledge memory, improving GPT-4o by 18.4% and boosting MOKA robotic manipulation success."},{"id":"arxiv:2406.02347","kind":"paper","n":379,"label":"Flash Diffusion: Accelerating Any Conditional Diffusion Model for Few Steps Image Generation","authors":"Clement Chadebec, Onur Tasar, Eyal Benaroche, Benjamin Aubin","year":2024,"url":"https://arxiv.org/abs/2406.02347","region":"t32","hemi":"wm","lobe":"frontal","color":"#8b4f9a","indegree":2,"alsoIn":[],"x":0.3935,"y":0.1115,"text":"Flash Diffusion is a distillation method that accelerates pretrained conditional diffusion models to generate high-quality images in 1-4 sampling steps. A LoRA student is trained to predict in a single step the multi-step (CFG-guided) teacher denoising of a partially-noised latent, combined with a latent-space adversarial (LSGAN) loss and a Distribution Matching Distillation (DMD) loss, using a custom mixture-of-Gaussians timestep distribution that over-samples the few inference timesteps. It reaches SOTA few-step FID/CLIP on COCO2014/2017 while training only 26.4M parameters in ~26 H100 GPU-hours, and generalizes across UNet (SD1.5/SDXL), DiT (Pixart-\u03b1), and MMDiT (SD3) backbones plus tasks like inpainting, super-resolution, face-swapping, and adapters."},{"id":"arxiv:2106.08261","kind":"paper","n":380,"label":"Physion: Evaluating Physical Prediction from Vision in Humans and Machines","authors":"Bear et al.","year":2021,"url":"https://arxiv.org/abs/2106.08261","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":4,"alsoIn":[],"x":0.3274,"y":0.6686,"text":"The paper introduces Physion, a dataset and benchmark for evaluating whether vision algorithms can predict how physical scenes evolve, using a unified object contact prediction (OCP) task \u2014 predicting whether a cued 'agent' object will touch a cued 'patient' object \u2014 across eight diverse physical scenarios (Dominoes, Support, Collide, Contain, Drop, Link, Roll, Drape) simulated in the ThreeDWorld (TDW) engine. The authors collected human judgments (800 participants, 100 per scenario) on the same 150 test stimuli per scenario and benchmarked four model classes spanning unsupervised pixel/object-centric predictors, supervised object models, ImageNet-pretrained encoders with learned dynamics, and particle-graph neural networks fed ground-truth physical state. They find vision models fall far short of humans while object-centric/ImageNet-pretrained encoders do better than pixel-based ones, and that particle-based graph networks with direct access to physical state approach human accuracy and produce more human-like error patterns. The conclusion is that extracting physical (object/particle) representations from vision is the main bottleneck to human-level and human-like physical prediction."},{"id":"arxiv:2512.04513","kind":"paper","n":381,"label":"BiTAgent: A Task-Aware Modular Framework for Bidirectional Coupling between Multimodal Large Language Models and World Models","authors":"","year":null,"url":"https://arxiv.org/abs/2512.04513","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":0,"alsoIn":[],"x":0.7617,"y":0.6613,"text":"BiTAgent addresses how to couple multimodal large language models (MLLMs) with world models (WMs) for embodied agents, noting prior connectors (GenRL, FOUNDER) are unidirectional and task-agnostic. It proposes a task-aware dynamic joint framework with bidirectional coupling: a forward path injecting MLLM semantics into the WM latent space via Task-Aware Modular Fusion (TAMF), and a backward path where WM-generated dense text-conditioned rewards refine the MLLM via joint optimization. On the DeepMind Control Suite across four embodiments, BiTAgent achieves the best overall normalized reward (0.91) and wins 9 of 10 tasks against model-free and model-based baselines. Ablations confirm each component (MLLM loss, TAMF, joint optimization) incrementally improves performance."},{"id":"arxiv:2109.06780","kind":"paper","n":382,"label":"Benchmarking the Spectrum of Agent Capabilities (Crafter)","authors":"Hafner","year":2022,"url":"https://arxiv.org/abs/2109.06780","region":"t22","hemi":"llm","lobe":"temporal","color":"#4f849a","indegree":2,"alsoIn":[],"x":0.639,"y":0.9066,"text":"The paper introduces Crafter, a procedurally-generated open-world 2D survival game (inspired by Minecraft) with 64x64x3 visual inputs that evaluates a broad spectrum of agent abilities within a single environment via 22 semantically meaningful achievements. Agents are trained for a 1M-step budget either with the provided sparse reward or purely from intrinsic objectives, and scored by the geometric mean of per-achievement success rates. Baselines show top RL methods reach only ~10% (DreamerV2) versus 50.5% for human experts, establishing Crafter as a challenging benchmark, and DreamerV2 trained longer exhibits emergent behaviors like tunnels, bridges, shelters, and plantations."},{"id":"arxiv:2407.18003","kind":"paper","n":383,"label":"Keep the Cost Down: A Review on Methods to Optimize LLM's KV-Cache Consumption","authors":"Luohe Shi, Hongyi Zhang, Yao Yao, et al.","year":2024,"url":"https://arxiv.org/abs/2407.18003","region":"t18","hemi":"llm","lobe":"temporal","color":"#67b09a","indegree":3,"alsoIn":[],"x":0.8993,"y":0.4474,"text":"A review of methods to optimize the KV Cache memory consumption of Transformer-based LLMs, which grows linearly with sequence length and bottlenecks long-context inference. It organizes techniques chronologically across three stages: training-stage architectural changes (MQA, GQA, cross-layer reuse, MLA), deploy-stage inference-system optimizations (PagedAttention/vLLM, distributed and offloaded caches), and post-training methods (eviction, merging, quantization). It also surveys long-context evaluation datasets and efficiency/capability metrics. The paper is a taxonomy and synthesis rather than a new method, with an accompanying GitHub paper list."},{"id":"arxiv:2503.20020","kind":"paper","n":384,"label":"Gemini Robotics: Bringing AI into the Physical World","authors":"Gemini Robotics Team, Google DeepMind","year":2025,"url":"https://arxiv.org/abs/2503.20020","region":"t0","hemi":"wm","lobe":"frontal","color":"#9a4f4f","indegree":1,"alsoIn":[],"x":0.1414,"y":0.7271,"text":"Google DeepMind introduces the Gemini Robotics family built on Gemini 2.0 to bring multimodal reasoning into physical robot control. It presents Gemini Robotics-ER, a VLM with enhanced embodied reasoning (2D/3D detection, pointing, grasp/trajectory prediction, multi-view correspondence), and Gemini Robotics, a Vision-Language-Action model that directly controls robots with a cloud backbone (<160ms) plus on-robot decoder achieving ~250ms end-to-end latency and 50Hz control. The models solve diverse dexterous manipulation tasks out of the box, follow open-vocabulary instructions, generalize across visual/instruction/action variations, and can be specialized to long-horizon dexterous tasks, learn new tasks from \u2264100 demos, and adapt to new embodiments (bi-arm Franka, Apollo humanoid). The paper also introduces the ERQA benchmark and addresses semantic action safety via constitutions and the ASIMOV datasets."},{"id":"arxiv:2406.06973","kind":"paper","n":385,"label":"RWKV-CLIP: A Robust Vision-Language Representation Learner","authors":"","year":null,"url":"https://arxiv.org/abs/2406.06973","region":"t23","hemi":"wm","lobe":"occipital","color":"#4f799a","indegree":7,"alsoIn":[],"x":0.3488,"y":0.3657,"text":"CLIP-style vision-language pre-training suffers from noisy web image-text pairs and the quadratic complexity of Transformer encoders. This paper contributes (1) a diverse description generation framework that fine-tunes LLaMA3-8B (distilled from ChatGPT instructions) to merge raw web text, OFA synthetic captions, and RAM++ open-set detection tags into cleaner descriptions, and (2) RWKV-CLIP, the first RWKV-driven dual-tower vision-language model combining Transformer-style parallel training with RNN-style efficient linear-complexity inference. Trained on YFCC15M and LAION10M/30M subsets, RWKV-CLIP reaches state-of-the-art on linear probing, zero-shot classification, and zero-shot image-text retrieval while using fewer parameters and FLOPs than CLIP-ViT."},{"id":"arxiv:2106.02039","kind":"paper","n":386,"label":"Offline Reinforcement Learning as One Big Sequence Modeling Problem (Trajectory Transformer)","authors":"Janner, Li, Levine","year":2021,"url":"https://arxiv.org/abs/2106.02039","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":4,"alsoIn":[],"x":0.1647,"y":0.241,"text":"The paper reframes offline reinforcement learning as a single sequence modeling problem, training a GPT-style Transformer decoder (Trajectory Transformer, TT) on autoregressively discretized sequences of states, actions, rewards, and rewards-to-go. Planning is done by repurposing beam search from NLP as a trajectory optimizer, biasing samples toward high cumulative reward. This unified approach dispenses with actor-critic components, explicit dynamics models, and pessimism/behavior-constraint machinery, yet matches or beats specialized offline RL methods on D4RL, and combined with a Q-function achieves state-of-the-art on sparse-reward AntMaze tasks."},{"id":"arxiv:2105.05233","kind":"paper","n":387,"label":"Diffusion Models Beat GANs on Image Synthesis","authors":"Prafulla Dhariwal, Alexander Nichol","year":2021,"url":"https://arxiv.org/abs/2105.05233","region":"t32","hemi":"wm","lobe":"frontal","color":"#8b4f9a","indegree":12,"alsoIn":[],"x":0.4259,"y":0.1515,"text":"The paper closes the gap between diffusion models and GANs on image synthesis through two contributions: a series of architecture ablations on the UNet (more attention heads/resolutions, BigGAN residual blocks, adaptive group normalization) that substantially boost FID, and classifier guidance, which uses gradients of a noise-aware classifier to steer sampling and trade diversity for fidelity via a gradient-scale hyperparameter. With these, the ablated diffusion model (ADM) plus guidance (ADM-G) beats BigGAN-deep on ImageNet 128/256/512 while maintaining higher recall (distribution coverage). Combining guidance with an upsampling stack yields the best results, and guided models match BigGAN with as few as 25 sampling steps."},{"id":"arxiv:2412.00887","kind":"paper","n":388,"label":"Playable Game Generation (PlayGen)","authors":"Yang et al.","year":2024,"url":"https://arxiv.org/abs/2412.00887","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":3,"alsoIn":[],"x":0.273,"y":0.5015,"text":"PlayGen tackles 'playable' game generation \u2014 real-time interaction, sufficient visual quality, and accurate simulation of interactive mechanics \u2014 which prior methods (Genie, MarioVGG, GameNGen) fail to deliver simultaneously. It combines (1) diverse agent-based game-data collection with cluster-based balanced sampling and self-supervised long-tailed transition learning, (2) an autoregressive DiT latent-diffusion model with an RNN-like hidden state (diffusion forcing) for long memory, and (3) a playability evaluation framework using action-aware ActAcc and ProbDiff metrics computed by a Valid Action Model. Validated on Super Mario Bros (2D) and Doom (3D), it sustains 20 FPS real-time play on a consumer RTX 2060 GPU and maintains accurate mechanics even after 1000+ frames."},{"id":"arxiv:2603.07670","kind":"paper","n":389,"label":"Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers","authors":"","year":2026,"url":"https://arxiv.org/abs/2603.07670","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":5,"alsoIn":[],"x":0.6182,"y":0.9583,"text":"This survey systematically reviews how memory is designed, implemented, and evaluated in LLM-based agents from 2022 through early 2026. It formalizes agent memory as a write\u2013manage\u2013read loop within a POMDP-style agent cycle, introduces a three-dimensional taxonomy (temporal scope, representational substrate, control policy), and examines five mechanism families: context-resident compression, retrieval-augmented stores, reflective self-improvement, hierarchical virtual context, and policy-learned management. It traces the evaluation shift from static recall benchmarks to multi-session agentic tests, surveys memory-critical application domains, and outlines open challenges including continual consolidation, causally grounded retrieval, trustworthy reflection, learned forgetting, and multimodal embodied memory."},{"id":"arxiv:2605.16689","kind":"paper","n":390,"label":"Against the Monolithic Wireless World Model: Why NextG Needs Composable and Agentic Intelligence","authors":"","year":null,"url":"https://arxiv.org/abs/2605.16689","region":"t39","hemi":"wm","lobe":"occipital","color":"#9a4f5b","indegree":3,"alsoIn":[],"x":0.2004,"y":0.5024,"text":"This position paper argues that the analogy between LLMs and wireless foundation models / wireless world models (WWMs) is structurally incomplete, because wireless data (CSI tensors, IQ samples, scheduler logs) lacks the broad, reusable, self-contained tokenized substrate that makes LLMs work. It identifies four structural data bottlenecks \u2014 configuration dependence, simulation/setup dependence, no universal wireless token, and missing operational feedback \u2014 that undermine monolithic models including MoE and WWMs. It concludes that the realistic near-term path to AI-native 6G/NextG is not a monolithic model but a composable, agentic architecture in which reasoning agents orchestrate specialized signal-processing models, classical algorithms, digital twins, standards-aware retrieval, and safety checks through explicit programmable interfaces. The paper also proposes 'wireless data cards' for documentation and sketches a three-layer agent-harness architecture with example workflows."},{"id":"arxiv:2506.09042","kind":"paper","n":391,"label":"Cosmos-Drive-Dreams: Scalable Synthetic Driving Data Generation with World Foundation Models","authors":"NVIDIA (Xuanchi Ren, Yifan Lu et al.)","year":2025,"url":"https://arxiv.org/abs/2506.09042","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":2,"alsoIn":[],"x":0.4399,"y":0.9494,"text":"Collecting rare, safety-critical driving data is costly, so the authors introduce Cosmos-Drive-Dreams, a synthetic data generation (SDG) pipeline powered by Cosmos-Drive \u2014 a suite of 7B world-foundation-model derivatives post-trained from NVIDIA Cosmos-1 for controllable, multi-view, spatiotemporally consistent driving video (and LiDAR) generation. The pipeline renders HDMap condition videos, uses an LLM prompt rewriter to diversify weather/time, generates and expands videos to six views, then applies a VLM rejection-sampling filter. Across 3D lane detection, camera and LiDAR 3D object detection, and driving-policy learning, the generated data consistently improves downstream performance, with the largest gains on long-tail corner cases like rain, fog, and night. The pipeline, models, and a synthetic dataset are open-sourced via NVIDIA's Cosmos platform."},{"id":"arxiv:2202.01682","kind":"paper","n":392,"label":"How to build a cognitive map","authors":"James C.R. Whittington, David McCaffary, Jacob J.W. Bakermans, Timothy E.J. Behr","year":2022,"url":"https://arxiv.org/abs/2202.01682","region":"t28","hemi":"llm","lobe":"frontal","color":"#7667b0","indegree":9,"alsoIn":[],"x":0.5658,"y":0.4881,"text":"This Perspective unifies diverse computational models of the hippocampal-entorhinal formation (successor representation, default representation, clone-structured cognitive graph, Tolman-Eichenbaum Machine, spatial memory pipeline, continuous attractor networks) into a common language for how the brain builds cognitive maps. It distils shared principles \u2014 representing latent states from sensory sequences, path integration as a compressed generalisable code, factorised/compositional bases, and hippocampal memories that link abstract structure to sensory content \u2014 and shows these account for both spatial and non-spatial neural representations. The authors offer novel (re)interpretations of phenomena (e.g. splitter/lap cells as latent states, representational drift as temporal remapping, grid-cell warping as a generalisation trade-off) and extend the framework to prefrontal task representations, credit assignment, replay, and higher cognition such as language and mathematics. It is a theoretical synthesis with illustrative TEM simulations rather than an empirical results paper."},{"id":"arxiv:2505.11329","kind":"paper","n":393,"label":"TokenWeave: Efficient Compute-Communication Overlap for Distributed LLM Inference","authors":"","year":null,"url":"https://arxiv.org/abs/2505.11329","region":"t24","hemi":"llm","lobe":"frontal","color":"#6784b0","indegree":1,"alsoIn":[],"x":0.8398,"y":0.3981,"text":"Distributed tensor-parallel LLM inference incurs 9\u201323% communication overhead even over NVLink, and existing compute-communication overlap techniques fail at the small token batches (\u22642K) used in low-latency serving because decomposition overheads overwhelm the gains. TokenWeave introduces a novel fused AllReduce\u2013RMSNorm kernel (using NVSHARP/Multimem with only 2\u20138 SMs) plus a wave-aware two-way token split that overlaps the communication of one split with the compute of the other, enabled even for batches as small as 1024 tokens. Integrated into vLLM-V1, it delivers up to 1.28\u00d7 latency speedup and up to 1.19\u00d7 throughput improvement across Llama-3.3-70B, Qwen2.5-72B and Mixtral-8x22B on 8\u00d7H100. In several settings it even beats a counterfactual baseline with all communication removed, because it also optimizes the memory-bound RMSNorm operation."},{"id":"oasis-model.github.io:525053739707","kind":"paper","n":394,"label":"Oasis: A Universe in a Transformer (real-time AI-generated Minecraft)","authors":"Decart & Etched","year":2024,"url":"https://oasis-model.github.io/","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":0,"alsoIn":[],"x":0.3823,"y":0.705,"text":"Oasis is a real-time, interactive, open-world AI model that generates a playable Minecraft-like experience entirely from a foundation model, taking user keyboard input and producing video that includes physics, game rules, and graphics with no underlying physics engine. The architecture pairs a ViT-based spatial autoencoder with a DiT-based latent diffusion backbone, generating frames autoregressively (each conditioned on user input) and trained with Diffusion Forcing. To maintain temporal stability over long horizons, it introduces dynamic noising that injects noise in early diffusion passes to curb error accumulation and removes it in later passes to persist high-frequency detail. Running on Decart's inference stack it achieves real-time interactivity, and the authors release code plus a 500M-parameter checkpoint, positioning the model for Etched's Sohu Transformer ASIC for future 4K, 100B+ scale."},{"id":"arxiv:2006.15134","kind":"paper","n":395,"label":"Critic Regularized Regression (CRR)","authors":"Wang et al.","year":2020,"url":"https://arxiv.org/abs/2006.15134","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":4,"alsoIn":[],"x":0.1788,"y":0.2923,"text":"The paper addresses offline (batch) RL, where standard off-policy algorithms fail due to overly optimistic Q-estimates and extrapolation beyond the fixed dataset. It proposes Critic Regularized Regression (CRR), which reduces offline policy optimization to value-filtered behavior cloning: actions in the dataset are weighted (binary indicator or exponential of an advantage estimate) by a learned distributional critic, plus a test-time Critic Weighted Policy (CWP) improvement step. CRR requires minimal changes to standard actor-critic methods yet scales to high-dimensional state/action spaces (e.g., 56-DoF humanoid, vision-based manipulation) and outperforms BC, D4PG, BCQ, and ABM across 17 RL Unplugged tasks plus 4 manipulation datasets."},{"id":"arxiv:2502.18080","kind":"paper","n":396,"label":"Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning","authors":"","year":null,"url":"https://arxiv.org/abs/2502.18080","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":9,"alsoIn":[],"x":0.847,"y":0.8187,"text":"The paper investigates whether excessively scaling Chain-of-Thought (CoT) length hurts LLM math reasoning, finding that longer CoTs can actually degrade performance\u2014especially on easier tasks\u2014because longer reasoning paths contain proportionally more erroneous reasoning rounds. The authors show an optimal reasoning-effort length exists that varies by task difficulty, and trace the harm to training on excessive erroneous steps. They propose Thinking-Optimal Scaling (TOPS): train a 'tag' model on a small seed set with Low/Medium/High reasoning efforts, generate responses at each effort, then self-improve on the shortest correct response per problem. Built on Qwen2.5-32B-Instruct, TOPS outperforms distillation-based 32B o1-like models and reaches parity with teacher QwQ-32B-Preview."},{"id":"arxiv:2405.06211","kind":"paper","n":397,"label":"A Survey on RAG Meeting LLMs: Towards Retrieval-Augmented Large Language Models","authors":"Wenqi Fan et al.","year":2024,"url":"https://arxiv.org/abs/2405.06211","region":"t31","hemi":"llm","lobe":"occipital","color":"#9767b0","indegree":8,"alsoIn":[],"x":0.8514,"y":0.6214,"text":"This survey reviews Retrieval-Augmented Large Language Models (RA-LLMs), which combine external knowledge retrieval with LLM generation to address LLM limitations such as hallucination, lack of domain-specific knowledge, and out-of-date internal knowledge. It organizes existing research along three technical perspectives \u2014 architecture (retrieval, generation, augmentation), training strategies (training-free, independent, sequential, joint), and application areas \u2014 detailing the challenges and capabilities at each stage. The paper surveys retriever types, retrieval granularity, pre-/post-retrieval enhancement, retrieval necessity/frequency, and augmentation integration points, then catalogs applications across NLP, recommendation, software engineering, AI-for-science, and finance, closing with future directions including trustworthy, multilingual, and multimodal RA-LLMs."},{"id":"arxiv:1805.09176","kind":"paper","n":398,"label":"Why the brain knows more than we do: non-conscious representations and their role in the construction of conscious experience","authors":"","year":null,"url":"https://arxiv.org/abs/1805.09176","region":"t5","hemi":"llm","lobe":"parietal","color":"#b09e67","indegree":0,"alsoIn":[],"x":0.5628,"y":0.6315,"text":"This review paper surveys evidence that non-conscious brain representations influence conscious cognition, perception, and action, and proposes a neural model explaining how statistical learning in resonant circuits generates temporal activity traces of non-conscious representations. It argues that reentrant signaling, top-down matching, and statistical coincidence of these traces progressively consolidate temporal patterns that constitute the neural signatures of conscious experience across long-distance brain networks, independent of spatial cortical maps. The framework integrates Adaptive Resonance Theory with evidence from psychophysics, neuroimaging, electrophysiology, and clinical studies."},{"id":"arxiv:2509.04633","kind":"paper","n":399,"label":"The Physical Basis of Prediction: World Model Formation in Neural Organoids via an LLM-Generated Curriculum","authors":"","year":null,"url":"https://arxiv.org/abs/2509.04633","region":"t5","hemi":"wm","lobe":"parietal","color":"#b09e67","indegree":3,"alsoIn":[],"x":0.1848,"y":0.4991,"text":"This position/framework paper proposes treating human neural organoids as embodied agents that learn internal world models, trained via a curriculum of three closed-loop virtual environments delivered through multi-electrode arrays (MEAs). It formalizes sensory encoding, motor decoding, and a free-energy-principle-based feedback scheme where predictable low-entropy stimuli act as reward and unpredictable high-entropy stimuli act as punishment to drive synaptic plasticity (LTP/LTD). A key methodological contribution is using an LLM as a meta-controller to automatically generate and optimize experimental protocols (JSON parameter tuning or full Python scripts). It also proposes a multi-scale evaluation that measures the physical correlates of the learned world model via electrophysiology (fEPSP slope), calcium imaging, and molecular markers. No empirical results are reported \u2014 it is a proposed framework with designs, pseudocode, and prompting examples."},{"id":"arxiv:2404.16767","kind":"paper","n":400,"label":"REBEL: Reinforcement Learning via Regressing Relative Rewards","authors":"Gao et al.","year":2024,"url":"https://arxiv.org/abs/2404.16767","region":"t26","hemi":"llm","lobe":"temporal","color":"#4f579a","indegree":3,"alsoIn":[],"x":0.66,"y":0.8385,"text":"PPO has become the default RL algorithm for fine-tuning generative models but requires value networks, clipping, and careful implementation, storing up to four billion-parameter models in memory. The paper proposes REBEL, which reduces policy optimization to a sequence of least-squares regression problems that predict the relative reward difference between two completions to a prompt in terms of the policy, eliminating value functions and clipping. Theoretically, REBEL generalizes Natural Policy Gradient (recovering it via a Gauss-Newton step) and achieves agnostic-setting regret bounds of O(1/sqrt(T) + sqrt(C_mu*epsilon)); it also extends to offline/hybrid data and intransitive preferences via the SPO reduction. Empirically REBEL matches or beats PPO, DPO, REINFORCE and RLOO on TL;DR summarization, general chat (Llama-3-8B), and text-to-image consistency-model tuning, while being simpler and more compute/memory efficient than PPO."},{"id":"deepmind.google:f020e8dd5471","kind":"paper","n":401,"label":"Genie 3: A New Frontier for World Models","authors":"Google DeepMind","year":2025,"url":"https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":3,"alsoIn":[],"x":0.3326,"y":0.6696,"text":"Genie 3 is a general-purpose foundation world model from Google DeepMind that generates diverse, interactive 3D environments from a text prompt, navigable in real time. Unlike prior Genie models, it is the first to combine real-time interaction with improved long-horizon consistency and realism, generating frames auto-regressively while attending to the growing past trajectory. It introduces 'promptable world events' that let users alter weather, objects, or characters mid-simulation, and was tested as a training/evaluation environment for the SIMA generalist embodied agent. It is released as a limited research preview to a small cohort of academics and creators."},{"id":"arxiv:2407.06508","kind":"paper","n":402,"label":"A Clinical Benchmark of Public Self-Supervised Pathology Foundation Models","authors":"Gabriele Campanella, Shengjia Chen, Ruchika Verma, Jennifer Zeng, et al.","year":2024,"url":"https://arxiv.org/abs/2407.06508","region":"t13","hemi":"wm","lobe":"parietal","color":"#6bb067","indegree":0,"alsoIn":[],"x":0.2911,"y":0.344,"text":"The paper introduces a clinical benchmark for evaluating publicly available self-supervised pathology foundation models on real-world, uncurated clinical whole-slide-image data from two institutions (Mount Sinai Health System and Memorial Sloan Kettering). It assembles 20 clinically relevant slide-level tasks spanning disease detection (9 cancer/IBD cohorts) and computational biomarker prediction (11 IHC/NGS/response tasks), and evaluates encoders (CTransPath, UNI, Virchow, Prov-GigaPath, an ImageNet ResNet50 baseline, and two in-house DINO ViTs SP21M/SP85M) using a frozen encoder plus a Gated MIL Attention aggregator. Key finding: DINO/DINOv2 models perform comparably and clearly beat ImageNet and CTransPath, but performance does not scale with model size, dataset size, or compute as in NLP/vision; pretraining dataset composition matters more, especially for biomarkers. The authors conclude current SSL strategies yield only incremental gains and may be approaching a limit."},{"id":"arxiv:2410.02664","kind":"paper","n":403,"label":"Grounded Answers for Multi-agent Decision-making Problem through Generative World Model","authors":"","year":null,"url":"https://arxiv.org/abs/2410.02664","region":"t22","hemi":"wm","lobe":"temporal","color":"#4f849a","indegree":6,"alsoIn":[],"x":0.1641,"y":0.4652,"text":"Generative models like GPT-4 produce sketchy, misleading answers for complex multi-agent decision-making problems because they lack trial-and-error grounding. The paper proposes Learning before Interaction (LBI), which integrates a language-guided world model\u2014a dynamics model (VQ-VAE image tokenizer + causal transformer generating transitions autoregressively) and a separate reward model (bidirectional transformer trained via inverse RL under task-description guidance)\u2014into the MARL pipeline to train a joint policy entirely in simulation and output an image-sequence answer. It also introduces VisionSMAC, paired state-image-language datasets built by a parser for the StarCraft Multi-Agent Challenge. LBI substantially outperforms imitation, offline-MARL, and online baselines on training and unseen SMAC maps, and produces consistent long-horizon rollouts and explainable rewards."},{"id":"arxiv:2403.02622","kind":"paper","n":404,"label":"World Models for Autonomous Driving: An Initial Survey","authors":"Guan et al.","year":2024,"url":"https://arxiv.org/abs/2403.02622","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":15,"alsoIn":[],"x":0.3507,"y":0.7894,"text":"An initial survey reviewing the state and prospects of world models in autonomous driving, covering their theoretical foundations (control theory, mental models, Pearl's causal hierarchy), architectural components, and applications. It explains core architectures\u2014the Recurrent State Space Model (RSSM) of the Dreamer series and LeCun's Joint-Embedding Predictive Architecture (JEPA)\u2014and catalogs ~40 world models across gaming, robotics, and video generation. For autonomous driving specifically, it organizes work into two areas: driving-scenario generation (GAIA-1, DriveDreamer, ADriver-I, WorldDreamer, MUVO, OccWorld) and planning/control (MILE, SEM2, Drive-WM, UniWorld, TrafficBots), then discusses challenges in long-term memory, sim-to-real generalization, and ethics/safety."},{"id":"arxiv:2502.20694","kind":"paper","n":405,"label":"WorldModelBench: Judging Video Generation Models As World Models","authors":"Li et al.","year":2025,"url":"https://arxiv.org/abs/2502.20694","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":7,"alsoIn":[],"x":0.4129,"y":0.8223,"text":"Existing video-generation benchmarks evaluate only general video quality (temporal consistency, aesthetics) and fail to assess whether models obey real-world dynamics, undermining claims that they are usable 'world models' for decision-making. WorldModelBench introduces a benchmark of 350 image+text condition pairs across 7 application-driven domains and 56 subdomains, scoring generated videos on instruction following, commonsense, and physics adherence (five physical laws) for both T2V and I2V models. The authors crowdsource 67K human labels (8336 complete votes) over 14 frontier models and fine-tune a 2B-parameter VILA judger that predicts human violation labels more accurately than GPT-4o. They further show that maximizing rewards from the fine-tuned judger via a reward-gradient (VADER-style) method improves a video model's world-modeling capability."},{"id":"arxiv:1711.00832","kind":"paper","n":406,"label":"A Unified Game-Theoretic Approach to Multiagent Reinforcement Learning (PSRO)","authors":"Marc Lanctot, Vinicius Zambaldi, Audrunas Gruslys, et al.","year":2017,"url":"https://arxiv.org/abs/1711.00832","region":"t22","hemi":"llm","lobe":"temporal","color":"#4f849a","indegree":4,"alsoIn":[],"x":0.6031,"y":0.8465,"text":"The paper tackles overfitting in multiagent reinforcement learning: policies learned by independent RL (InRL) overfit to co-training partners and fail to generalize at execution. It introduces Joint Policy Correlation (JPC) to quantify this effect, and proposes Policy-Space Response Oracles (PSRO) \u2014 a general MARL algorithm that uses deep RL to compute approximate best responses to mixtures (meta-strategies) of opponent policies and empirical game-theoretic analysis to update those meta-strategies, generalizing InRL, iterated best response, double oracle, and fictitious play. A scalable parallel variant, Deep Cognitive Hierarchies (DCH), with decoupled (bandit) meta-solvers reduces memory to O(n\u00b2K\u00b2). Experiments in partially-observable gridworld coordination games and Leduc poker show PSRO/DCH sharply reduce JPC and yield robust counter-strategies that safely exploit opponents."},{"id":"arxiv:2010.02193","kind":"paper","n":407,"label":"Mastering Atari with Discrete World Models (DreamerV2)","authors":"Hafner et al.","year":2021,"url":"https://arxiv.org/abs/2010.02193","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":12,"alsoIn":[],"x":0.1693,"y":0.3227,"text":"Modeling Atari games accurately enough to derive successful behaviors had remained an open challenge, with world models failing to compete with state-of-the-art model-free agents on competitive benchmarks. The paper introduces DreamerV2, which learns behaviors purely by imagining trajectories in the compact latent space of a separately trained world model whose stochastic state is a vector of categorical (discrete) variables, trained with KL balancing. DreamerV2 is the first agent to reach human-level performance on the 55-game Atari benchmark by learning entirely within a learned world model, surpassing the final performance of top single-GPU model-free agents IQN and Rainbow using the same compute and wall-clock budget. The same agent also solves humanoid stand-up and walking from pixel inputs in continuous-control settings."},{"id":"arxiv:2504.19516","kind":"paper","n":408,"label":"Boosting LLM Serving through Spatial-Temporal GPU Resource Sharing","authors":"","year":null,"url":"https://arxiv.org/abs/2504.19516","region":"t24","hemi":"llm","lobe":"frontal","color":"#6784b0","indegree":5,"alsoIn":[],"x":0.9157,"y":0.6083,"text":"LLM serving suffers from low GPU utilization because the compute-intensive prefill phase and memory-bound decode phase mismatch, and chunked-prefill hybrid batching trades off throughput against latency while wasting resources. The authors diagnose two root causes (wave-quantization/attention bottlenecks in prefill, and chunked-prefill's biased latency-over-throughput tradeoff) and present Bullet, a spatial-temporal orchestration system that concurrently executes prefill and decode requests with fine-grained, SLO-aware SM provisioning driven by an analytical performance model and lightweight CUDA-stream SM masking. On real-world workloads Bullet delivers 1.26x average (up to 1.55x) throughput gains over state-of-the-art serving systems while meeting latency SLOs and raising SM utilization to 86.2% (11.2% higher than SGLang)."},{"id":"arxiv:2312.14125","kind":"paper","n":409,"label":"VideoPoet: A Large Language Model for Zero-Shot Video Generation","authors":"Dan Kondratyuk et al. (Google)","year":2023,"url":"https://arxiv.org/abs/2312.14125","region":"t29","hemi":"llm","lobe":"parietal","color":"#6a4f9a","indegree":5,"alsoIn":[],"x":0.6839,"y":0.1797,"text":"VideoPoet is a decoder-only autoregressive transformer that treats video generation as a language modeling problem, encoding images, video, and audio into discrete tokens (via MAGVIT-v2 and SoundStream tokenizers) within a unified ~300,000-entry vocabulary alongside T5 XL text embeddings. It follows an LLM-style two-stage protocol \u2014 multimodal multi-task pretraining (text-to-video, image-to-video, frame prediction, inpainting/outpainting, stylization, audio-video) followed by task-specific adaptation \u2014 yielding a single model that handles many video tasks zero-shot. The 8B model achieves state-of-the-art zero-shot text-to-video results (e.g., MSR-VTT FVD 213, UCF-101 FVD 355) and is consistently preferred by human raters over diffusion baselines on motion interestingness and realism, demonstrating that LLMs can rival diffusion for high-fidelity, large-motion video generation."},{"id":"arxiv:2311.13549","kind":"paper","n":410,"label":"ADriver-I: A General World Model for Autonomous Driving","authors":"Jia, Mao et al.","year":2023,"url":"https://arxiv.org/abs/2311.13549","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":1,"alsoIn":[],"x":0.33,"y":0.8004,"text":"ADriver-I is a general world model for autonomous driving that unifies low-level control-signal prediction and future-scene generation by combining a multimodal LLM (Vicuna-7B-1.5 + CLIP-ViT-Large) with a video latent diffusion model (built on Stable Diffusion 2.1). It introduces the 'interleaved vision-action pair' to encode visual tokens and control signals (speed, steer angle) in a shared text-embedding space, letting the MLLM autoregressively predict the current control signal, which then conditions the VDM to generate future frames; feeding generated frames back enables recurrent 'infinite driving' in a self-created world. On nuScenes with three historical vision-action pairs, it reaches L1 errors of 0.072 m/s (speed) and 0.091 rad (steer) and FID 5.5 / FVD 97 for four predicted future frames, outperforming MLP/CNN/ViT baselines and prior generators without needing 3D boxes or HD maps."},{"id":"arxiv:2509.20328","kind":"paper","n":411,"label":"Video models are zero-shot learners and reasoners","authors":"Wiedemer, Li, Vicol et al. (Google DeepMind)","year":2025,"url":"https://arxiv.org/abs/2509.20328","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":2,"alsoIn":[],"x":0.4025,"y":0.6234,"text":"The paper investigates whether large generative video models can serve as general-purpose vision foundation models the way LLMs became foundation models for language. The authors prompt Google DeepMind's Veo 3 (and predecessor Veo 2) zero-shot \u2014 with only an input image plus text instruction \u2014 across 18,384 generated videos spanning 62 qualitative and 7 quantitative tasks organized into a perception \u2192 modeling \u2192 manipulation \u2192 reasoning hierarchy. Veo 3 solves tasks it was never trained for (segmentation, edge detection, image editing, intuitive physics, affordances, maze/symmetry/analogy solving) and shows 'chain-of-frames' visual reasoning, with large, consistent gains from Veo 2 to Veo 3. Bespoke task-specific models still win on absolute performance, but the rapid Veo 2\u2192Veo 3 jump and no-plateau pass@10 scaling suggest video models are on a trajectory toward unified vision foundation models."},{"id":"philarchive.org:69d8d2ad404e","kind":"paper","n":412,"label":"In Search of a Biological Crux for AI Consciousness","authors":"Bradford Saad","year":2025,"url":"https://philarchive.org/archive/SAAISOv1","region":"t2","hemi":"llm","lobe":"temporal","color":"#b07d67","indegree":2,"alsoIn":[],"x":0.6017,"y":0.7326,"text":"The provided document does not contain the actual paper; the retrieval was blocked by a Cloudflare security page from philarchive.org, so no body text, methods, experiments, or results are available. Based solely on the title, 'In Search of a Biological Crux for AI Consciousness' appears to be a philosophy/neuroscience essay examining whether there is a specifically biological prerequisite ('crux') that would determine whether artificial systems could be conscious. No abstract, argument, or findings could be extracted from the supplied text."},{"id":"nature.com:413d00acfd43","kind":"paper","n":413,"label":"Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects","authors":"Rao, R. P. N. & Ballard, D. H.","year":1999,"url":"https://www.nature.com/articles/nn0199_79","region":"t28","hemi":"llm","lobe":"frontal","color":"#7667b0","indegree":0,"alsoIn":[],"x":0.5972,"y":0.5421,"text":"The paper proposes a hierarchical predictive-coding model of the visual cortex in which feedback connections from higher to lower cortical areas carry predictions of lower-level neural activity, while feedforward connections carry only the residual error between predictions and actual activity. A hierarchical network of model neurons trained on natural images developed simple-cell-like receptive fields, and the error-coding neurons reproduced endstopping and other extra-classical (nonclassical surround) receptive-field effects. The authors argue these surround effects need not be purely feedforward but can arise from cortico-cortical feedback as a by-product of efficiently encoding natural-image statistics."},{"id":"arxiv:2510.04978","kind":"paper","n":414,"label":"Aligning Perception, Reasoning, Modeling and Interaction: A Survey on Physical AI","authors":"Kun Xiang, Terry Jingchen Zhang, Yinya Huang, et al.","year":2026,"url":"https://arxiv.org/abs/2510.04978","region":"t7","hemi":"llm","lobe":"occipital","color":"#adb067","indegree":9,"alsoIn":[],"x":0.6914,"y":0.5207,"text":"This survey proposes a unified framework for 'Physical AI' that organizes the field into four progressive, interdependent capabilities\u2014physical perception, physics reasoning, world modeling, and embodied interaction\u2014arguing that physical understanding emerges cumulatively from passive observation toward active intervention. It introduces a three-tier, output-centric capability\u2013task\u2013subtask taxonomy, formalizes Physical AI as a partially observed controlled dynamical system, and reviews 300+ papers spanning architectures, training, causal inference, and embodied systems. Drawing on benchmark evidence, it shows frontier models still rely on statistical pattern matching rather than internalized causal/physical law, consistently underperforming humans on physical reasoning, dynamics prediction, and interaction tasks. The authors document the survey construction with a PRISMA-style pipeline and conclude that hybrid physics-grounded, neuro-symbolic, and embodied approaches are the most promising path forward."},{"id":"arxiv:2605.22882","kind":"paper","n":415,"label":"GEM-4D: Geometry-Enhanced Video World Models for Robot Manipulation","authors":"","year":null,"url":"https://arxiv.org/abs/2605.22882","region":"t15","hemi":"wm","lobe":"occipital","color":"#4f9a62","indegree":2,"alsoIn":[],"x":0.4181,"y":0.6372,"text":"Video world models generate plausible futures but fail to track the same physical points consistently across time, breaking action extraction for robot manipulation. GEM-4D injects dense 4D correspondence supervision distilled from a frozen pretrained geometry foundation model into a video diffusion transformer during training via a parallel geometry flow-matching branch, forcing the video backbone's internal features to encode depth, camera pose, and scene motion without altering its output space or adding inference cost. An adaptive inverse dynamics module (AIDS) converts the correspondence-consistent rollouts into executable 6-DoF trajectories. GEM-4D reaches SOTA on 4D video prediction and geometric consistency and raises real-world Droid manipulation success from 61% to 81%."},{"id":"arxiv:2308.07870","kind":"paper","n":416,"label":"A Survey on Brain-Inspired Deep Learning via Predictive Coding","authors":"Tommaso Salvatori, Ankur Mali, Christopher L. Buckley, Thomas Lukasiewicz, Rajes","year":2023,"url":"https://arxiv.org/abs/2308.07870","region":"t20","hemi":"llm","lobe":"frontal","color":"#67b0b0","indegree":10,"alsoIn":[],"x":0.6166,"y":0.6416,"text":"This survey reviews predictive coding (PC), a brain-inspired learning theory grounded in variational inference, as an alternative to error backpropagation for training deep neural networks. It formalizes PC as inversion of hierarchical Gaussian generative models via free-energy minimization using local, Hebbian-style updates, and traces its history from 1950s signal compression through Rao & Ballard's visual-cortex model to the modern free-energy formulation. It surveys PC applications across supervised learning, NLP, computer vision, temporal/continual learning, associative memory, and active inference/control, alongside the neural generative coding (NGC) generalization, software frameworks, and neuromorphic hardware. It concludes that PC offers locality, parallelism, robustness, and biological plausibility but currently faces scalability and efficiency bottlenecks that prevent it from matching backprop on large-scale tasks."},{"id":"philarchive.org:04bcc5cbd425","kind":"paper","n":417,"label":"Testing for Consciousness in Current AI","authors":"Patrick Butlin","year":2024,"url":"https://philarchive.org/archive/BUTTFC","region":"t2","hemi":"llm","lobe":"temporal","color":"#b07d67","indegree":3,"alsoIn":[],"x":0.6031,"y":0.7297,"text":"The provided text is not a paper but a Cloudflare security block page preventing access to philarchive.org. No paper content, methods, results, or findings could be extracted. The title 'Testing for Consciousness in Current AI' suggests a theoretical/evaluative work on AI consciousness, but no substantive content is available."},{"id":"x.com:351692e95fe5","kind":"paper","n":418,"label":"On generation vs. causal prediction from a world model (public statement)","authors":"Yann LeCun","year":2024,"url":"https://x.com/ylecun/status/1758740106955952191","region":"t6","hemi":"wm","lobe":"temporal","color":"#b0a967","indegree":2,"alsoIn":[],"x":0.2963,"y":0.4764,"text":"In a public statement, Yann LeCun argues that producing realistic-looking videos from prompts does not demonstrate that a system understands the physical world, because generation only requires sampling one plausible video whereas causal prediction of the (much smaller) space of plausible continuations of a real video \u2014 especially when conditioned on an action \u2014 is far harder. He contends that generating full pixel-level continuations is both expensive and pointless, and that it is preferable to predict abstract representations that discard action-irrelevant scene detail. This is the rationale for JEPA (Joint Embedding Predictive Architecture), which is non-generative and predicts in representation space. He cites VICReg, I-JEPA, and V-JEPA as evidence that joint-embedding architectures learn better visual representations than generative pixel-reconstruction methods."},{"id":"arxiv:1012.5151","kind":"paper","n":419,"label":"Statistical theory of isotropic turbulence Part IV: multiscales and cascade","authors":"","year":null,"url":"https://arxiv.org/abs/1012.5151","region":"t36","hemi":"llm","lobe":"frontal","color":"#b06793","indegree":3,"alsoIn":[],"x":0.66,"y":0.5273,"text":"This paper analyzes the multiscale and cascade aspects of isotropic turbulence using an explicit map method derived from a new Sedov-type solution of the turbulence scaling equation. The author shows that the energy cascading process in isotropic turbulence is governed by a logistic map, producing an infinite sequence of period-doubling bifurcations that constitutes the first deductive evidence of the Richardson cascade. The bifurcation parameter depends only on the relative Reynolds number, and the K41 power-law spectrum k^(-5/3) is recovered as an asymptotic state of the exact solution."},{"id":"arxiv:2402.09664","kind":"paper","n":420,"label":"CodeMind: Evaluating Large Language Models for Code Reasoning","authors":"Liu et al.","year":2024,"url":"https://arxiv.org/abs/2402.09664","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":3,"alsoIn":[],"x":0.8843,"y":0.7604,"text":"The paper introduces CodeMind, a framework to evaluate the code reasoning abilities of LLMs through three tasks: Independent Execution Reasoning (IER, explicit output prediction), Specification Reasoning (SR, implicitly incorporating test data into code synthesis), and Dynamic Semantics Reasoning (DSR, refactoring to a shorter semantically-equivalent program). Evaluating 13 LLMs on 1450 Python programs from HumanEval, CRUXEval, ClassEval, and Avatar, it finds that models can reason about some dynamic code aspects but degrade sharply with higher complexity, nested constructs, non-primitive types, and intra-class dependencies. It shows the three tasks evaluate LLMs differently (so all are needed) and that bug-repair success is not correlated with code reasoning except for frontier reasoning models \u2014 implying many repairs succeed via natural-language shortcuts, hallucinations, or code clones rather than genuine understanding."},{"id":"arxiv:2504.05352","kind":"paper","n":421,"label":"Achieving binary weight and activation for LLMs using Post-Training Quantization","authors":"","year":null,"url":"https://arxiv.org/abs/2504.05352","region":"t37","hemi":"llm","lobe":"parietal","color":"#b06788","indegree":5,"alsoIn":[],"x":0.8858,"y":0.367,"text":"The paper tackles the performance collapse of LLM post-training quantization (PTQ) below 4 bits for both weights and activations. It proposes a W(1+1)A(1\u00d74) framework (storage-equivalent to W2A4) that binarizes weights to 1 bit plus 1 bit of fine-grained group affiliation, and decomposes INT4 activations into four INT1 channels, so the FC inner-loop becomes pure Boolean (XOR/AND/popcount) operations. Weights are quantized with an EM-based, Hessian-aware fine-grained grouping scheme and activations are refined by smoothing the four per-bit scaling factors to cancel first-order error. On Wikitext2 it reaches perplexity 8.58 (LLaMA-7B) and 8.89 (LLaMA2-7B) at W2A4, far surpassing prior SOTA and approaching FP16 (5.68 / 5.47)."},{"id":"arxiv:1906.10184","kind":"paper","n":422,"label":"A free energy principle for a particular physics","authors":"Karl Friston","year":2019,"url":"https://arxiv.org/abs/1906.10184","region":"t25","hemi":"llm","lobe":"parietal","color":"#6779b0","indegree":4,"alsoIn":[],"x":0.5717,"y":0.8108,"text":"This monograph develops a 'free energy principle for a particular physics' \u2014 a theory of what it means to be a 'thing' defined by a Markov blanket (statistical separation of internal from external states) that persists at nonequilibrium steady-state (NESS). Starting from Langevin/Fokker-Planck dynamics and a Helmholtz decomposition of flow into dissipative (gradient) and solenoidal components, Friston shows that quantum mechanics (Schr\u00f6dinger equation), statistical mechanics (fluctuation theorems, Jarzynski equality) and classical mechanics (Hamiltonian/Lagrangian, Maxwell) all emerge as limiting cases governed by the amplitude of random fluctuations. For autonomous 'active' particles he derives a Bayesian mechanics in which internal states parameterise a variational density that infers external states, so self-organisation is recast as variational/active inference (self-evidencing) via minimisation of particular and expected free energy. The arguments are illustrated throughout with numerical simulations of a synthetic 'primordial soup' of coupled Lorenz systems that self-organises into a virus/Bacillus-like particle with a Markov blanket."},{"id":"arxiv:2210.02747","kind":"paper","n":423,"label":"Flow Matching for Generative Modeling","authors":"Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, Matt Le","year":2022,"url":"https://arxiv.org/abs/2210.02747","region":"t3","hemi":"wm","lobe":"occipital","color":"#9a714f","indegree":5,"alsoIn":[],"x":0.317,"y":0.1288,"text":"The paper introduces Flow Matching (FM), a simulation-free objective for training Continuous Normalizing Flows by regressing a neural vector field onto the vector field of a fixed conditional probability path, avoiding the expensive ODE simulations that previously made CNF training intractable at scale. It proves (via Conditional Flow Matching, CFM) that a per-sample objective has identical gradients to the intractable marginal objective, and generalizes Gaussian probability paths to subsume diffusion (VE/VP) as special cases while enabling new Optimal-Transport (straight-line) conditional paths. On ImageNet/CIFAR-10 it shows FM with OT paths beats diffusion baselines on likelihood, FID, and sampling cost (number of function evaluations). The OT paths give straight, constant-speed sampling trajectories that train faster and sample more efficiently than curved diffusion paths."},{"id":"arxiv:2403.04121","kind":"paper","n":424,"label":"Can Large Language Models Reason and Plan?","authors":"Kambhampati","year":2024,"url":"https://arxiv.org/abs/2403.04121","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":11,"alsoIn":[],"x":0.7539,"y":0.8145,"text":"A position/critique paper arguing that LLMs do not perform principled reasoning or planning but rather 'universal approximate retrieval' over web-scale corpora, akin to a non-veridical external System 1. The author summarizes empirical studies (PlanBench) showing GPT-3/3.5/4 fail at autonomous plan generation, that performance collapses under name obfuscation, and that LLM self-verification makes results worse. It proposes the LLM-Modulo framework, where LLMs generate candidate solutions/knowledge that are checked by sound external model-based verifiers (or expert humans), to gainfully leverage LLMs without ascribing reasoning to them."},{"id":"arxiv:2105.07308","kind":"paper","n":425,"label":"Towards a Predictive Processing Implementation of the Common Model of Cognition","authors":"","year":null,"url":"https://arxiv.org/abs/2105.07308","region":"t5","hemi":"llm","lobe":"parietal","color":"#b09e67","indegree":2,"alsoIn":[],"x":0.5702,"y":0.5809,"text":"This position/architecture paper proposes a neural implementation of the Common Model of Cognition (the consensus across ACT-R, Soar, and SIGMA) built from two biologically plausible building blocks: Neural Generative Coding (NGC), a predictive-processing model, and holographic (vector-symbolic) memory. NGC circuits are mapped onto the perceptual sensory cortices, motor cortex (via active NGC for action), and basal ganglia (action selection/task switching via competitive Hebbian learning), while holographic vectors implement the working-memory buffers and long-term declarative memory. The authors argue this combination yields scalable local Hebbian update rules that resist catastrophic interference and support continual, single-trial, and transfer learning. The paper is a conceptual proposal/blueprint with a planned validation agenda rather than reported empirical results."},{"id":"arxiv:2603.18532","kind":"paper","n":426,"label":"Scaling Sim-to-Real Reinforcement Learning for Robot VLAs with Generative 3D Worlds","authors":"","year":null,"url":"https://arxiv.org/abs/2603.18532","region":"t0","hemi":"wm","lobe":"frontal","color":"#9a4f4f","indegree":9,"alsoIn":[],"x":0.1123,"y":0.6826,"text":"The paper addresses the loss of generality when RL fine-tuning robot vision-language-action (VLA) models in the narrowly-scoped real world, which overfits broadly pretrained policies to specific scenes. The authors use 3D world generative models (EmbodiedGen) plus a GPT-4o-powered language-driven scene designer to automatically generate hundreds of diverse interactive 3D scenes in the ManiSkill 3 simulator, then fine-tune a \u03c00 flow-matching VLA across them with massively parallel RL using a new PPOFlow algorithm. Starting from a pretrained imitation baseline, the method raises simulation pick-and-place success from 9.7% to 79.8% and real-world success from 21.7% to 75%, with task-completion speedups. Ablations show zero-shot generalization scales positively with the number of training scenes."},{"id":"arxiv:2402.01528","kind":"paper","n":427,"label":"Decoding Speculative Decoding","authors":"","year":null,"url":"https://arxiv.org/abs/2402.01528","region":"t9","hemi":"llm","lobe":"parietal","color":"#97b067","indegree":6,"alsoIn":[],"x":0.8773,"y":0.769,"text":"This paper performs a large-scale empirical study (over 352 experiments) of speculative decoding for LLM inference, using LLAMA-65B and OPT-66B as target models and OPT/LLAMA-family models as draft models. The authors find that the dominant bottleneck is draft-model latency (driven by model depth), and that a draft model's language-modeling accuracy does not correlate strongly with its token acceptance rate (TAR). Based on these insights, they prune draft models with Sheared-LLAMA to trade depth for width (keeping parameter count constant), producing shallow-wide drafts that cut latency while preserving TAR. Their NoFT-Wide-796M draft delivers up to 111% higher throughput than the existing Sheared-LLAMA-1.3B draft and generalizes across LLAMA 1/2/3.1 and supervised fine-tuned target models."},{"id":"arxiv:2501.18837","kind":"paper","n":428,"label":"Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming","authors":"Sharma et al.","year":2025,"url":"https://arxiv.org/abs/2501.18837","region":"t8","hemi":"llm","lobe":"frontal","color":"#a2b067","indegree":2,"alsoIn":[],"x":0.8055,"y":0.6743,"text":"The paper addresses the vulnerability of LLMs to universal jailbreaks that systematically bypass safeguards to extract harmful CBRN information. It introduces Constitutional Classifiers\u2014input and output classifier safeguards trained on synthetic data generated from natural-language rules (a constitution) specifying permitted and restricted content. Across 3,000+ hours of human red teaming, no universal jailbreak was found, and on automated evaluations the classifiers blocked 95% of held-out jailbreak attempts versus 14% without classifiers, with only 0.38% absolute increase in production refusals and 23.7% inference overhead."},{"id":"arxiv:2109.01134","kind":"paper","n":429,"label":"Learning to Prompt for Vision-Language Models","authors":"","year":null,"url":"https://arxiv.org/abs/2109.01134","region":"t23","hemi":"wm","lobe":"occipital","color":"#4f799a","indegree":8,"alsoIn":[],"x":0.2973,"y":0.3643,"text":"Deploying CLIP-like vision-language models on downstream image recognition requires laborious prompt engineering, where slight wording changes drastically alter accuracy. The paper proposes Context Optimization (CoOp), which replaces hand-crafted prompt context words with continuous learnable vectors trained via cross-entropy while keeping the entire pre-trained CLIP backbone frozen, offering unified-context and class-specific-context (CSC) variants. Across 11 image classification datasets, CoOp beats hand-crafted prompts with as few as one or two shots, averages ~15% gain at 16 shots (over 45% on EuroSAT), outperforms the linear-probe baseline, and is more robust to domain shift than zero-shot CLIP."},{"id":"arxiv:2506.21539","kind":"paper","n":430,"label":"WorldVLA: Towards Autoregressive Action World Model","authors":"Cen, Yu, Yuan, Jiang, Huang et al. (Alibaba DAMO Academy)","year":2025,"url":"https://arxiv.org/abs/2506.21539","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":2,"alsoIn":[],"x":0.2011,"y":0.762,"text":"WorldVLA is an autoregressive model that unifies a Vision-Language-Action (VLA) policy and a world model in a single framework built on the Chameleon backbone, encoding images, text, and actions into a shared discrete vocabulary. The world model predicts future frames from current images and actions (learning environment physics) while the action model generates actions from observations, and the two components mutually improve each other. The authors also identify that naive autoregressive generation of action chunks causes error propagation, and propose an action attention masking strategy that blocks the current action from attending to prior actions. On the LIBERO benchmark, WorldVLA beats the same-backbone discrete action model and improves video generation, and the attention mask largely recovers the performance lost during action-chunk generation."},{"id":"arxiv:2510.19818","kind":"paper","n":431,"label":"Semantic World Models","authors":"Berg, Zhu, Bao, Durugkar et al.","year":2025,"url":"https://arxiv.org/abs/2510.19818","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":4,"alsoIn":[],"x":0.2185,"y":0.8315,"text":"Conventional world models predict future pixels for planning, but strong pixel reconstruction does not reliably correlate with good planning decisions. This paper reframes world modeling as a visual-question-answering problem about task-relevant semantic information in future states, finetuning a pretrained VLM (PaliGemma 3B) on image-action-question-answer data to answer questions about the consequences of an action sequence. The resulting Semantic World Model (SWM) is combined with sampling-based (MPPI) and gradient-based planning to improve a base policy, raising average success from 14.4% to 81.6% on LangTable and 45.33% to 76.0% on OGBench while outperforming pixel-based video-diffusion and offline-RL baselines and generalizing to out-of-distribution scenes."},{"id":"arxiv:2411.13112","kind":"paper","n":432,"label":"DriveMLLM: A Benchmark for Spatial Understanding with Multimodal Large Language Models in Autonomous Driving","authors":"Xianda Guo, et al.","year":2024,"url":"https://arxiv.org/abs/2411.13112","region":"t12","hemi":"wm","lobe":"frontal","color":"#5e9a4f","indegree":3,"alsoIn":[],"x":0.3092,"y":0.8938,"text":"The paper introduces SURDS, the first large-scale benchmark for evaluating fine-grained spatial reasoning of vision-language models in autonomous-driving scenes, built on nuScenes with 41,080 training and 9,250 validation VQA instances across six spatial tasks (yaw orientation, pixel localization, depth range, pairwise distance, left/right ordering, front/behind). Benchmarking frontier VLMs (GPT-4o, Gemini, Qwen) exposes persistent failures in absolute localization and multi-object relational reasoning, with scale not predicting competence. The authors propose a post-training pipeline that adds SFT on auto-generated chain-of-thought data followed by GRPO alignment with custom location and logic process rewards (plus accuracy and format rewards). Their Qwen2.5-VL-3B-SFT-GRPO-LocLogic variant reaches an overall score of 40.80, beating GPT-4o (13.30), Gemini-2.0-flash (35.71), and Qwen2.5-VL-72B (33.47)."},{"id":"arxiv:1706.03762","kind":"paper","n":433,"label":"Attention Is All You Need","authors":"Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N","year":2017,"url":"https://arxiv.org/abs/1706.03762","region":"t16","hemi":"llm","lobe":"frontal","color":"#67b084","indegree":4,"alsoIn":[],"x":0.8051,"y":0.3347,"text":"The paper addresses the sequential-computation bottleneck of RNN/CNN-based sequence transduction models that limits parallelization and long-range dependency learning. It introduces the Transformer, an encoder-decoder architecture built entirely on (multi-head, scaled dot-product) self-attention, dispensing with recurrence and convolution. On WMT 2014 machine translation it sets new state of the art \u2014 28.4 BLEU EN-DE and 41.8 BLEU EN-FR \u2014 while training far faster and cheaper, and generalizes to English constituency parsing."},{"id":"arxiv:2307.15818","kind":"paper","n":434,"label":"RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control","authors":"Brohan, Brown, Carbajal, Zitkovich et al. (Google DeepMind)","year":2023,"url":"https://arxiv.org/abs/2307.15818","region":"t0","hemi":"wm","lobe":"frontal","color":"#9a4f4f","indegree":5,"alsoIn":[],"x":0.1667,"y":0.7862,"text":"RT-2 addresses how to transfer Internet-scale vision-language knowledge directly into low-level robotic control rather than using VLMs only for high-level planning. The authors co-fine-tune pretrained VLMs (PaLI-X 5B/55B and PaLM-E 12B) jointly on web vision-language data and robot trajectories, representing 6-DoF end-effector actions as discretized text tokens so a single model outputs actions in the same way it outputs language. Across ~6,000 real-world evaluation trials, RT-2 roughly doubles generalization success over RT-1/MOO baselines (62% vs 32%/35% unseen average) and shows emergent symbol-understanding, reasoning, and human-recognition skills absent from robot data. Adding chain-of-thought 'Plan' steps further enables multi-stage semantic reasoning."},{"id":"arxiv:2212.13472","kind":"paper","n":435,"label":"Optimal scheduling of island integrated energy systems considering multi-uncertainties and hydrothermal simultaneous transmission: A deep reinforcement learning approach","authors":"","year":null,"url":"https://arxiv.org/abs/2212.13472","region":"t4","hemi":"wm","lobe":"frontal","color":"#9a7c4f","indegree":5,"alsoIn":[],"x":0.3067,"y":0.1463,"text":"The paper tackles real-time optimal scheduling of an island integrated energy system (IES) facing multi-uncertainties from renewable power sources and electricity/heat/freshwater loads. It models a novel island IES (CHP, CWP, GT, GB, WT units) with a seawater-desalination-based combined water-and-power unit and a proposed 'hydrothermal simultaneous transmission' (HST) structure that carries heat and freshwater in one pipeline, then casts scheduling as an MDP solved by a model-free Distributed Proximal Policy Optimization (DPPO) deep RL agent. Tested on real data from a North China island, DPPO converges after ~10,000 episodes and achieves the lowest daily operating cost (\u00a5139,770.95) and fastest decision time (0.0137 s) versus SAC, WOA, PSO, and the interior point method. The trained network also adjusts dispatch in real time to simulated emergencies (sudden load/wind changes) without manual intervention or forecasting."},{"id":"nature.com:e173b62796c2","kind":"paper","n":436,"label":"Abstract representations emerge in human hippocampal neurons during inference","authors":"Hristos S. Courellis, Juri Minxha, Araceli R. Cardenas, et al., Ueli Rutishauser","year":2024,"url":"https://www.nature.com/articles/s41586-024-07799-x","region":"t28","hemi":"llm","lobe":"frontal","color":"#7667b0","indegree":2,"alsoIn":[],"x":0.5538,"y":0.4995,"text":"This study investigates how abstract, generalizable neural representations emerge in the human brain during inferential reasoning. Single-unit activity was recorded from 2,694 neurons across hippocampus, amygdala, medial frontal cortex (preSMA, dACC, vmPFC) and ventral temporal cortex in 17 epilepsy patients (36 included sessions) performing a serial reversal-learning task with an uncued latent context. Using representational-geometry metrics (CCGP, parallelism score, shattering dimensionality) over all 35 balanced dichotomies of 8 task conditions, the authors find that only the hippocampus simultaneously encodes latent context and stimulus pairing in an abstract (disentangled) format, and only in sessions where patients successfully performed inference. This abstract geometry appears whether inference was learned by trial-and-error or installed within minutes by verbal instruction, and its presence tracks correct vs. error trials, linking representational format to behaviour."},{"id":"arxiv:2505.19867","kind":"paper","n":437,"label":"Deep Active Inference Agents for Delayed and Long-Horizon Environments","authors":"Yavar Taheri Yeganeh, Mohsen Jafari, Andrea Matta","year":2025,"url":"https://arxiv.org/abs/2505.19867","region":"t14","hemi":"llm","lobe":"temporal","color":"#67b06e","indegree":5,"alsoIn":[],"x":0.5792,"y":0.8928,"text":"The paper introduces Deep Active Inference (DAIF) agents that fold an explicit, differentiable policy network inside a single generative (world) model trained under the active-inference formalism, so the model can predict an entire H-step horizon in one look-ahead and policy optimization proceeds by back-propagating expected-free-energy (EFE) gradients rather than exhaustive tree search. Training alternates between calibrating the generative model via variational free energy (VFE) and updating the actor via EFE gradients, using an experience replay buffer; at deployment a single gradient step every H steps replaces per-step planning. Evaluated on a high-fidelity simulation of an energy-efficiency control problem for parallel automotive manufacturing machines (delayed, long-horizon, highly stochastic), DAIF reaches 2.59%\u00b10.16% production loss with 12.49%\u00b10.04% energy saving, beating DQN baselines, and raises energy efficiency per production unit by 10.21%\u00b10.14% with negligible throughput loss over a one-month simulation. It remains effective even with very long overshooting horizons (H up to 1000) and works without handcrafted rewards."},{"id":"openreview.net:3f48a33ff9bf","kind":"paper","n":438,"label":"A Path Towards Autonomous Machine Intelligence","authors":"Yann LeCun","year":2022,"url":"https://openreview.net/pdf?id=BZ5a1r-kVsf","region":"t14","hemi":"llm","lobe":"temporal","color":"#67b06e","indegree":5,"alsoIn":[],"x":0.6779,"y":0.642,"text":"This position paper by Yann LeCun proposes a modular, fully-differentiable cognitive architecture for building autonomous intelligent agents that learn world models by observation, much like humans and animals. The architecture comprises six modules \u2014 configurator, perception, world model, cost (intrinsic cost + trainable critic), short-term memory, and actor \u2014 operating in a reactive 'Mode-1' (System 1) or a deliberative model-predictive 'Mode-2' (System 2) planning loop. Its centerpiece is the Joint Embedding Predictive Architecture (JEPA) and its hierarchical extension (H-JEPA), a non-generative world model trained with non-contrastive self-supervised learning that predicts in abstract representation space rather than pixel space, using regularized latent variables to handle uncertainty and enable hierarchical planning. The paper is conceptual: it argues against pure scaling, pure reinforcement learning ('reward is enough'), and hard-wired symbol manipulation, and proposes energy-based, regularized (non-contrastive) learning as the path forward, but reports no experimental results."},{"id":"arxiv:2503.16419","kind":"paper","n":439,"label":"Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models","authors":"Yang Sui et al.","year":2025,"url":"https://arxiv.org/abs/2503.16419","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":8,"alsoIn":[],"x":0.8291,"y":0.8207,"text":"This is the first structured survey on efficient reasoning for Large Language Models, addressing the 'overthinking phenomenon' where Large Reasoning Models (e.g., OpenAI o1, DeepSeek-R1) generate verbose, redundant Chain-of-Thought sequences that inflate inference cost and latency. It organizes existing work into three families: model-based efficient reasoning (RL with length-based rewards, SFT on variable-length CoT), reasoning output-based efficient reasoning (compressing steps into latent representations, dynamic reasoning paradigms during inference), and input prompts-based efficient reasoning (prompt-guided length control and difficulty-based routing). It additionally covers efficient training data, reasoning in small language models via distillation/compression, and evaluation/benchmarking, and maintains a public living repository of the literature."},{"id":"arxiv:2602.08024","kind":"paper","n":440,"label":"FlashVID: Efficient Video Large Language Models via Training-free Tree-based Spatiotemporal Token Merging","authors":"","year":null,"url":"https://arxiv.org/abs/2602.08024","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":1,"alsoIn":[],"x":0.3611,"y":0.5188,"text":"VLLMs must process large numbers of visual tokens, making video understanding computationally expensive, and existing acceleration methods compress spatial and temporal redundancy independently\u2014ignoring that semantically correlated visual features shift in position, scale, and orientation over time. FlashVID is a training-free, plug-and-play inference acceleration framework with two modules: Attention and Diversity-based Token Selection (ADTS), which picks informative and diverse tokens per frame via a calibrated Max-Min Diversity Problem, and Tree-based Spatiotemporal Token Merging (TSTM), which jointly models spatial and temporal redundancy through hierarchical spatiotemporal redundancy trees. Across three VLLMs (LLaVA-OneVision, LLaVA-Video, Qwen2.5-VL) and five video benchmarks, FlashVID retains 99.1% of LLaVA-OneVision's accuracy while pruning 90% of visual tokens and outperforms prior SOTA (FastV, VisionZip, PruneVID, FastVID). Used as a token-budget extender, it lets Qwen2.5-VL ingest 10\u00d7 more frames for an 8.6% relative gain at the same compute."},{"id":"arxiv:2410.10377","kind":"paper","n":441,"label":"Learning Sub-Second Routing Optimization in Computer Networks requires Packet-Level Dynamics (PackeRL)","authors":"Patrick Kr\u00e4mer, Andreas Blenk et al.","year":2024,"url":"https://arxiv.org/abs/2410.10377","region":"t39","hemi":"wm","lobe":"occipital","color":"#9a4f5b","indegree":1,"alsoIn":[],"x":0.2948,"y":0.1867,"text":"The paper argues that learning sub-second routing optimization (RO) in computer networks requires packet-level network dynamics rather than the fluid/flow-based models commonly used, because packet interactions (especially TCP sending-rate behavior) shape true traffic dynamics. The authors build PackeRL, the first packet-level RL training/evaluation environment for routing in arbitrary generic topologies (backed by the ns-3 simulator and a scenario generator, synnet), and show that MAGNNETO \u2014 an RL policy trained in a fluid environment \u2014 fails to beat static shortest-path baselines when moved to packet-level conditions. They introduce two GNN-based RL policies: M-Slim, a link-weight/shortest-path method that needs only one All-Pairs-Shortest-Path (APSP) pass per update, and FieldLines, a next-hop selection policy that re-optimizes any topology in milliseconds without re-training. Both outperform EIGRP/OSPF and MAGNNETO in high-volume and TCP-heavy traffic while generalizing from training on networks of \u226410 nodes up to 250-node topologies."},{"id":"arxiv:2405.08295","kind":"paper","n":442,"label":"SpeechVerse: A Large-scale Generalizable Audio Language Model","authors":"Nilaksh Das et al.","year":2024,"url":"https://arxiv.org/abs/2405.08295","region":"t11","hemi":"llm","lobe":"occipital","color":"#81b067","indegree":7,"alsoIn":[],"x":0.8574,"y":0.2652,"text":"SpeechVerse is a multi-task audio-language framework that connects a frozen self-supervised speech encoder (WavLM Large or BEST-RQ) to a frozen Flan-T5-XL LLM via a small learnable 1-D convolution downsampling module and LoRA adapters, trained with multimodal instruction finetuning and a two-stage curriculum. Using continuous latent speech representations and natural-language instructions, it performs 11 diverse speech tasks (ASR, 5 SLU, 5 paralinguistic) and generalizes zero-shot to unseen prompts and tasks. The multi-task model beats conventional task-specific baselines on 9 of 11 tasks, and constrained/joint decoding strategies further improve generalization on unseen tasks by up to 21% absolute."},{"id":"arxiv:2306.09265","kind":"paper","n":443,"label":"LVLM-eHub: A Comprehensive Evaluation Benchmark for Large Vision-Language Models","authors":"","year":null,"url":"https://arxiv.org/abs/2306.09265","region":"t30","hemi":"llm","lobe":"temporal","color":"#8c67b0","indegree":4,"alsoIn":[],"x":0.8305,"y":0.1956,"text":"The paper presents LVLM-eHub, a comprehensive evaluation benchmark for Large Vision-Language Models, evaluating 8 representative LVLMs (BLIP2, LLaVA, LLaMA-Adapter V2, MiniGPT-4, mPLUG-Owl, Otter, InstructBLIP, VPGTrans) across 6 capability categories on 47 benchmarks plus an online arena platform for human-judged pairwise battles. Key findings include: InstructBLIP overfits in-domain tasks (best on 5 quantitative categories) but generalizes poorly in the open-world arena; instruction-tuned LVLMs with moderate data suffer from object hallucination; and multi-turn reasoning evaluation can mitigate hallucination. The work reveals that current evaluation metrics like CIDEr are ineffective for diverse LVLM outputs."},{"id":"arxiv:2002.04501","kind":"paper","n":444,"label":"Some interesting observations on the free energy principle","authors":"","year":null,"url":"https://arxiv.org/abs/2002.04501","region":"t25","hemi":"llm","lobe":"parietal","color":"#6779b0","indegree":4,"alsoIn":[],"x":0.5867,"y":0.8106,"text":"This paper is a rebuttal by Friston, Da Costa, and Parr to Biehl et al.'s (2020) technical critique of the free energy principle (FEP) as formulated in Friston's 2013 'Life as we know it'. The authors concede that the critique's observations are interesting and some correct, but argue none invalidate the FEP; they use three of Biehl et al.'s observations to clarify the technical foundations of the FEP. Key clarifications concern which solenoidal coupling terms are precluded when a Markov blanket emerges under sparse coupling, why an unbounded evidence bound does not threaten the free energy lemma, and why the gradients of the evidence bound vanish on the internal manifold. The core conclusion is that self-organisation to nonequilibrium steady-state can always be read as (approximating) Bayesian inference regardless of the bound's magnitude."},{"id":"arxiv:2309.13638","kind":"paper","n":445,"label":"Embers of Autoregression: Understanding Large Language Models Through the Problem They Are Trained to Solve","authors":"McCoy, Yao, Friedman, Hardy & Griffiths","year":2023,"url":"https://arxiv.org/abs/2309.13638","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":8,"alsoIn":[],"x":0.8066,"y":0.7882,"text":"The paper argues that LLMs should be understood through the problem they were trained to solve\u2014next-word prediction over Internet text\u2014an approach the authors call the 'teleological approach.' From this they derive three hypothesized 'embers of autoregression': LLM accuracy is influenced by task probability, output probability, and input probability, even on deterministic tasks where probability should be irrelevant. Evaluating GPT-3.5 and GPT-4 on eleven tasks (shift ciphers, Pig Latin, acronyms, counting, reversal, sorting, linear functions, etc.), they find robust evidence for all three effects, with output probability mattering more than input probability. Surprising failure modes emerge\u2014e.g., GPT-4 decodes a cipher at 51% when the answer is high-probability text but only 13% when low-probability\u2014leading them to conclude LLMs should be evaluated as a distinct kind of system rather than as humans."},{"id":"arxiv:1903.11027","kind":"paper","n":446,"label":"nuScenes: A Multimodal Dataset for Autonomous Driving","authors":"Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qia","year":2020,"url":"https://arxiv.org/abs/1903.11027","region":"t12","hemi":"llm","lobe":"frontal","color":"#5e9a4f","indegree":0,"alsoIn":[],"x":0.6286,"y":0.1867,"text":"nuScenes is the first large-scale autonomous-driving dataset to carry the full sensor suite \u2014 6 cameras, 5 radars and 1 lidar with full 360\u00b0 coverage \u2014 comprising 1000 20s scenes fully annotated with 3D bounding boxes for 23 classes and 8 attributes, plus semantic maps and scene descriptions. It introduces new 3D detection and tracking metrics (NDS, true-positive error metrics, and recall-averaged tracking metrics) designed for the AV application, and provides lidar (PointPillars) and image (OFT, MonoDIS) baselines. The dataset is roughly 7x the annotations and 100x the images of KITTI, and the authors show that accumulating multiple lidar sweeps and using larger training volumes substantially improve detection."},{"id":"arxiv:2402.12275","kind":"paper","n":447,"label":"WorldCoder, a Model-Based LLM Agent: Building World Models by Writing Code and Interacting with the Environment","authors":"Tang, Key & Ellis","year":2024,"url":"https://arxiv.org/abs/2402.12275","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":1,"alsoIn":[],"x":0.7004,"y":0.8153,"text":"WorldCoder is a model-based LLM agent that synthesizes a Python program representing the world's transition and reward functions from its interactions with the environment, then plans over that learned program to act. The key contribution is a learning objective combining data-consistency (\u03d51) with 'optimism under uncertainty' (\u03d52) \u2014 a logical constraint between a program and a planner that drives goal-directed exploration \u2014 plus curriculum-based transfer by editing existing code. On Sokoban, Minigrid, and AlfWorld it is far more sample-efficient than deep RL, more compute-efficient than ReAct-style agents (O(1) vs O(T) LLM calls per task), and transfers knowledge across environments. The authors also prove polynomial sample complexity (D\u00d7(K+1) actions) for discovering a good-enough world model."},{"id":"arxiv:2410.07484","kind":"paper","n":448,"label":"WALL-E: World Alignment by Rule Learning Improves World Model-based LLM Agents","authors":"Zhou et al.","year":2024,"url":"https://arxiv.org/abs/2410.07484","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":1,"alsoIn":[],"x":0.6794,"y":0.8258,"text":"The paper tackles the gap between LLMs' prior knowledge and a deployed environment's true dynamics, which makes raw LLMs unreliable as world models for model-based agents. It proposes WALL-E, a neurosymbolic approach that performs gradient-free 'world alignment' by inducing, refining, translating-to-code, and pruning (via maximum set coverage) a small set of rules learned from comparing agent-explored trajectories against world-model predictions; the aligned LLM+rules world model then drives an LLM agent via model-predictive control (MPC). On Minecraft (MineDojo TechTree) and ALFWorld, WALL-E achieves higher success rates with fewer replanning rounds and lower token cost than prior LLM/VLM agents. It exceeds Minecraft baselines by 15-30% in success rate (avg 69% vs GITM 54%) while using 8-20 fewer replanning rounds and only 60-80% of tokens, and reaches a 95% ALFWorld success rate by the 6th iteration."},{"id":"arxiv:2001.10371","kind":"paper","n":449,"label":"Improving operational flexibility of integrated energy system with uncertain renewable generations considering thermal inertia of buildings","authors":"","year":null,"url":"https://arxiv.org/abs/2001.10371","region":"t4","hemi":"wm","lobe":"frontal","color":"#9a7c4f","indegree":2,"alsoIn":[],"x":0.2133,"y":0.1353,"text":"Addresses insufficient operational flexibility from traditional 'heat-set' CHP operation that curtails renewable energy in winter heating periods. Proposes a chance-constrained programming (CCP) optimal scheduling model for a small-scale integrated energy system (IES) with CHP units, thermal units, wind/PV, and auxiliary equipment (BESS, HST, EB), building a heating-load model from building thermal inertia and thermal comfort, and providing probabilistic spinning reserves. A sequence operation theory (SOT) approach converts the chance constraint into a deterministic equivalent, reformulating the model as MILP solved by CPLEX. Tested on a modified IEEE 30-bus system, the method fully absorbs renewable curtailment and minimizes generation cost while balancing economy and reliability."},{"id":"1x.tech:d48ab282348b","kind":"paper","n":450,"label":"1X World Model: Evaluating Bits, not Atoms","authors":"1X World Model Team","year":2024,"url":"https://www.1x.tech/1x-world-model.pdf","region":"t17","hemi":"llm","lobe":"parietal","color":"#4f9a79","indegree":0,"alsoIn":[],"x":0.6375,"y":0.4181,"text":"1X presents a generative video world model (1XWM) that predicts future robot observations and a learned task-level state value (success/progress) for a full-body humanoid, conditioned on exact low-level action trajectories rather than text prompts. The model acts as a fast offline policy evaluator: it forecasts contact-rich, full-body manipulation outcomes so different policies can be compared under identical initial conditions, replacing slow, expensive real-robot evaluation. The authors show that value-prediction alignment with real outcomes scales with data (especially autonomous on-policy rollouts and failure data), transfers across tasks, and correlates with real-world A/B evaluations, enabling checkpoint selection and architecture ablations. They report multi-task transfer lifting Shelf alignment from 63.06% to 71.17% when adding Arcade data."},{"id":"arxiv:2509.06266","kind":"paper","n":451,"label":"Spatial Reasoning with Vision-Language Models in Ego-Centric Multi-View Scenes","authors":"Mohsen Gholami, Ahmad Rezaei, Zhou Weimin, Yong Zhang, Mohammad Akbari","year":2025,"url":"https://arxiv.org/abs/2509.06266","region":"t12","hemi":"wm","lobe":"frontal","color":"#5e9a4f","indegree":7,"alsoIn":[],"x":0.2289,"y":0.908,"text":"Current VLMs struggle with 3D spatial reasoning, and existing benchmarks only cover single images or static indoor videos rather than the ego-centric multi-view inputs real embodied agents use. The authors introduce Ego3D-Bench, a benchmark of 8,600+ human-curated QA pairs across 5 categories built from nuScenes, Waymo, and Argoverse outdoor multi-view data, and benchmark 16 SOTA VLMs, revealing a large gap to human performance. They also propose Ego3D-VLM, a plug-and-play post-training framework that builds a textual cognitive map from estimated global 3D coordinates (via REC + metric depth), yielding ~12% average accuracy gain on multi-choice QA and ~56% RMSE improvement on absolute distance estimation. The method generalizes to All-Angle Bench and VSI-Bench and works with any existing VLM."},{"id":"arxiv:2410.11767","kind":"paper","n":452,"label":"Analyzing (In)Abilities of SAEs via Formal Languages","authors":"Abhinav Menon, Manish Shrivastava, David Krueger, Ekdeep Singh Lubana","year":2024,"url":"https://arxiv.org/abs/2410.11767","region":"t16","hemi":"llm","lobe":"frontal","color":"#67b084","indegree":5,"alsoIn":[],"x":0.6536,"y":0.3958,"text":"The paper stress-tests sparse autoencoders (SAEs) for language-model interpretability using synthetic formal-language testbeds (Dyck-2, Expr, and a simple English PCFG), training small 2-block Transformers on these grammars and SAEs on their first-block activations under many hyperparameter settings. Interpretable latents (depth/counter variables, part-of-speech features) reliably emerge\u2014especially with top-k rather than L1 regularization\u2014but feature identifiability is highly sensitive to inductive biases (normalization, regularization, pre-bias), and strongly correlational features turn out to have no causal effect on model computation under clamping interventions. The authors argue causality must be a first-class training target and propose a causal regularization loss that treats within-sequence token correlations as weak supervision via latent interpolation. This loss yields features with predictable causal function, but it is biased toward MLP-mediated (attention-independent) features, so it works for static-next-token parts of speech yet finds nothing for attention-driven Dyck-2/Expr."},{"id":"arxiv:2602.12520","kind":"paper","n":453,"label":"Multi-Agent Model-Based Reinforcement Learning with Joint State-Action Learned Embeddings","authors":"Vladimir Egorov, Aleksei Shpilman","year":2026,"url":"https://arxiv.org/abs/2602.12520","region":"t22","hemi":"llm","lobe":"temporal","color":"#4f849a","indegree":3,"alsoIn":[],"x":0.6914,"y":0.7968,"text":"The paper tackles sample inefficiency and coordination in cooperative multi-agent RL under partial observability by fusing model-based RL with joint state-action representation learning. It introduces MMSA, a framework that augments a VAE-based world model and a QMIX-style value-decomposition network with State-Action Learned Embeddings (SALE), using the world model to generate latent imagined roll-outs that feed a monotonic mixing network under the CTDE paradigm. Across MAMuJoCo, Level-Based Foraging, SMAC, and SMACv2, MMSA matches or exceeds model-free and model-based baselines (e.g., 0.93 mean win rate on SMACv1 vs 0.63 for MAMBA) while training faster, and ablations show each component (world model, SALE, KL balancing, global state) is necessary."},{"id":"arxiv:2606.15874","kind":"paper","n":454,"label":"LLM-as-Code: Agentic Programming for Agent Harness","authors":"","year":null,"url":"https://arxiv.org/abs/2606.15874","region":"t35","hemi":"llm","lobe":"occipital","color":"#b0679e","indegree":1,"alsoIn":[],"x":0.6555,"y":0.9207,"text":"This position paper argues that the dominant LLM-agent design \u2014 putting the LLM in charge of orchestration (the ReAct reason-and-act loop) \u2014 commits a category error by assigning deterministic control-flow work (looping, branching, sequencing, termination) to a probabilistic sampler, which structurally produces token explosion, control-flow hallucination, and unreliable completion that no better prompt or stronger model can fix. It proposes Agentic Programming / 'LLM-as-Code,' where ordinary program code owns all control flow and the LLM is invoked only at leaf nodes that genuinely need reasoning, with context built from the execution call tree as a DAG so each call's context length scales with call depth O(depth) rather than accumulated steps O(steps). A GUI computer-use agent built this way reaches 86.8% overall success on OSWorld in just 15 steps, beating the strongest 100-step baseline (Holo3-35B-A3B at 80.4%). The paradigm extends naturally to parallel multi-agent collaboration (agents as sibling call-graph branches) and self-programmed evolution (regenerated functions committed as durable code)."},{"id":"arxiv:2511.22904","kind":"paper","n":455,"label":"Language-conditioned world model improves policy generalization by reading environmental descriptions","authors":"Nguyen & Lee","year":2025,"url":"https://arxiv.org/abs/2511.22904","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":12,"alsoIn":[],"x":0.0993,"y":0.4959,"text":"The paper tackles the problem of agents understanding dynamics-descriptive language (how an environment behaves) rather than just task instructions, and generalizing to unseen games with novel dynamics and novel surface language. It proposes LED-WM (Language-aware Encoder for Dreamer World Model), built on DreamerV3, which adds a cross-modal attention encoder that explicitly grounds manual sentences to entity symbols in a grid observation; a policy is then trained purely from the world model without inference-time planning or expert demonstrations. On the MESSENGER and MESSENGER-WM benchmarks, LED-WM generalizes better than model-free and model-based baselines and beats EMMA-LWM in all compositional settings without expert data. The authors further show the trained world model can be used to fine-tune the policy on synthetic test trajectories for additional (though limited) gains."},{"id":"philpapers.org:e7fc6fd114b8","kind":"paper","n":456,"label":"Neurology and the Mind-Brain Problem","authors":"Roger W. Sperry","year":1952,"url":"https://philpapers.org/rec/SPENAT","region":"t2","hemi":"llm","lobe":"temporal","color":"#b07d67","indegree":7,"alsoIn":[],"x":0.6106,"y":0.7345,"text":"No paper content was retrievable: the supplied 'abstract' and 'paper text' are both a Cloudflare security block page from philpapers.org (HTTP access-denied notice, Ray ID a11b240c2c89fa36), not the actual article. Based solely on the title, 'Neurology and the Mind-Brain Problem' appears to be a theoretical/philosophical work on the relationship between neural processes and mind. No method, experiments, results, or numeric findings are available to index."},{"id":"arxiv:2402.08268","kind":"paper","n":457,"label":"World Model on Million-Length Video And Language With Blockwise RingAttention","authors":"Liu, Yan, Zaharia & Abbeel","year":2024,"url":"https://arxiv.org/abs/2402.08268","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":0,"alsoIn":[],"x":0.4235,"y":0.5225,"text":"This paper tackles long-context understanding in sequence models by training Large World Model (LWM), a family of 7B Llama-2-based models capable of processing text and video sequences exceeding 1M tokens. The method uses Blockwise RingAttention for exact attention without approximation, progressive context extension from 32K to 1M tokens (scaling RoPE \u03b8 alongside), model-generated question-answering data for chat, and a VQGAN tokenizer plus masked sequence packing to jointly train on interleaved text, image, and video tokens. LWM achieves near-perfect single-needle retrieval over its full 1M context, competitive multi-needle retrieval versus GPT-4, and best-in-class long-video understanding among 7B models on Video-MME. The work also open-sources the optimized training/inference implementation and the any-to-any model that can also generate images and video."},{"id":"arxiv:2601.17067","kind":"paper","n":458,"label":"A Mechanistic View on Video Generation as World Models: State and Dynamics","authors":"(authors listed on arXiv)","year":2026,"url":"https://arxiv.org/abs/2601.17067","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":9,"alsoIn":[],"x":0.4025,"y":0.6845,"text":"This survey reconciles the gap between modern 'stateless' large-scale video generation models (Sora, Veo, Kling, Wan, etc.) and classic state-centric world model theory from cognitive science, control theory, and model-based RL. It proposes a two-pillar taxonomy\u2014State Construction (implicit context-management memory vs. explicit latent compression) and Dynamics Modeling (causal knowledge integration with LMMs vs. causal architecture reformulation)\u2014to organize how video models can become world simulators. It argues evaluation must move from visual-fidelity metrics (IS/FID/FVD) to functional benchmarks of physical persistence and causal reasoning, and identifies enhancing persistence (data-driven memory, compressed fidelity) and advancing causality (latent factor decoupling, reasoning-prior integration) as the two key frontiers."},{"id":"arxiv:2402.06590","kind":"paper","n":459,"label":"Predictive representations: building blocks of intelligence","authors":"Wilka Carvalho, Momchil S. Tomov, William de Cothi, Caswell Barry, Samuel J. Ger","year":2024,"url":"https://arxiv.org/abs/2402.06590","region":"t5","hemi":"llm","lobe":"parietal","color":"#b09e67","indegree":7,"alsoIn":[],"x":0.5687,"y":0.5925,"text":"This survey argues that predictive representations\u2014specifically the successor representation (SR) and its generalizations, the successor model (SM) and successor features (SF)\u2014serve as versatile 'building blocks of intelligence' bridging model-free and model-based reinforcement learning. It formalizes each representation via its cumulant and Bellman/TD update (Table 1), reviews practical algorithms for scaling them (USFAs, GPI, Option Keyboard, contrastive/density-estimation SMs), and catalogs AI applications spanning exploration, transfer, hierarchical RL, jumpy model-based RL, and multi-agent coordination. It then marshals converging evidence from neuroscience (hippocampal place/grid cells as SR/eigenvectors, dopamine as vector-valued prediction error, replay) and cognitive science (revaluation, multi-task transfer, spatial navigation, memory) that biological brains use SR-like predictive representations. The core thesis is that caching predictions of future state occupancy resolves flexibility\u2013efficiency trade-offs across both artificial and biological intelligence."},{"id":"arxiv:2211.10831","kind":"paper","n":460,"label":"Joint Embedding Predictive Architectures Focus on Slow Features","authors":"Vlad Sobal, Jyothir S. V., Siddhartha Jalagam, Nicolas Carion, Kyunghyun Cho, Ya","year":2022,"url":"https://arxiv.org/abs/2211.10831","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":4,"alsoIn":[],"x":0.2419,"y":0.4069,"text":"The paper studies whether Joint Embedding Predictive Architectures (JEPA), trained reconstruction-free with VICReg or SimCLR objectives, can learn useful world models in a fully offline, reward-free setting, compared against pixel-reconstruction and inverse-dynamics baselines. Using a toy moving-dot environment with controllable background distractors (uniform/structured noise that is either changing or fixed per frame) and linear probing for the dot's location, they find JEPA matches or beats reconstruction when distractor noise changes every time step but fails when the noise is temporally fixed. They give a theoretical argument that JEPA objectives are minimized by a trivial solution that copies static 'slow' noise features and ignores the moving dot. They conclude JEPA focuses on slow features, a limitation they suggest addressing with hierarchical architectures or anti-constant-representation constraints."},{"id":"arxiv:2403.14403","kind":"paper","n":461,"label":"Adaptive-RAG: Learning to Adapt Retrieval-Augmented LLMs through Question Complexity","authors":"Soyeong Jeong, Jinheon Baek, Sukmin Cho, Sung Ju Hwang, Jong C. Park","year":2024,"url":"https://arxiv.org/abs/2403.14403","region":"t31","hemi":"llm","lobe":"occipital","color":"#9767b0","indegree":2,"alsoIn":[],"x":0.8664,"y":0.662,"text":"Retrieval-augmented LLMs for QA either waste compute on simple queries (multi-step retrieval) or fail on complex multi-hop queries (single/no retrieval). Adaptive-RAG trains a small classifier (T5-Large) to predict a query's complexity level (A: no retrieval, B: single-step, C: multi-step iterative) and routes each query to the matching strategy, using automatically collected training labels from model prediction outcomes plus dataset inductive biases (no human annotation). Across six open-domain QA datasets and three LLMs (FLAN-T5-XL/XXL, GPT-3.5-Turbo-Instruct), it improves accuracy-efficiency trade-off over adaptive baselines, e.g. averaged EM 37.17 / F1 46.94 with FLAN-T5-XL versus the multi-step approach's F1 48.85 at far lower cost. It substantially outperforms Adaptive Retrieval and Self-RAG among adaptive methods."},{"id":"arxiv:2411.04118","kind":"paper","n":462,"label":"Medical Adaptation of Large Language and Vision-Language Models: Are We Making Progress?","authors":"Daniel P. Jeong et al.","year":2024,"url":"https://arxiv.org/abs/2411.04118","region":"t13","hemi":"llm","lobe":"parietal","color":"#6bb067","indegree":2,"alsoIn":[],"x":0.8484,"y":0.2508,"text":"The paper investigates whether domain-adaptive pretraining (DAPT) on biomedical corpora actually improves medical LLMs/VLMs over their general-domain base models. Through apples-to-apples head-to-head comparisons of 7 medical LLMs and 2 medical VLMs against their exact base models on 13 textual and 8 visual QA datasets, with per-model prompt optimization and bootstrap statistical testing, the authors find that all medical VLMs and nearly all medical LLMs fail to consistently improve over their base models in zero-/few-shot QA. They show that common sloppy practices\u2014single fixed prompts optimized only for the medical model and ignoring statistical uncertainty\u2014produce misleadingly optimistic conclusions about medical DAPT."},{"id":"arxiv:2312.15796","kind":"paper","n":463,"label":"GenCast: Diffusion-based Ensemble Forecasting for Medium-Range Weather","authors":"Price et al.","year":2023,"url":"https://arxiv.org/abs/2312.15796","region":"t34","hemi":"wm","lobe":"temporal","color":"#9a4f93","indegree":6,"alsoIn":[],"x":0.2393,"y":0.1369,"text":"GenCast is a probabilistic machine-learning weather prediction (MLWP) model that produces global 15-day ensemble forecasts via a conditional diffusion model trained on 40 years of ERA5 reanalysis. It autoregressively samples 12-hour steps at 0.25\u00b0 resolution for 80+ surface/atmospheric variables, generating each 15-day forecast in ~8 minutes on a Cloud TPUv5. Evaluated against ECMWF's operational ensemble ENS, GenCast has greater CRPS skill on 97.4% of 1320 targets and better predicts extreme weather, tropical cyclone tracks, and regional wind power. The work shows generative diffusion models can capture high-dimensional weather distributions accurately enough to beat the top operational NWP ensemble."},{"id":"arxiv:2311.17918","kind":"paper","n":464,"label":"Driving into the Future: Multiview Visual Forecasting and Planning with World Model for Autonomous Driving","authors":"Wang et al. (Drive-WM)","year":2023,"url":"https://arxiv.org/abs/2311.17918","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":2,"alsoIn":[],"x":0.3689,"y":0.7947,"text":"The paper introduces Drive-WM, the first multiview world model for autonomous driving that is compatible with existing end-to-end planners. Built on latent video diffusion, it jointly models temporal and multiview dimensions and adds a view-factorization scheme (generating 'stitched' views conditioned on 'reference' views) plus a unified condition interface for images, text, 3D layouts, and ego actions. Drive-WM generates high-fidelity, controllable, consistent multiview driving videos and is used for tree-based planning that imagines multiple futures and selects trajectories via image-based rewards. On nuScenes it achieves leading video quality (15.8 FID / 122.7 FVD), strong multiview consistency (KPM up to 94.4%), and improves planner robustness in out-of-distribution cases."},{"id":"arxiv:1911.10601","kind":"paper","n":465,"label":"Scaling active inference","authors":"","year":null,"url":"https://arxiv.org/abs/1911.10601","region":"t14","hemi":"wm","lobe":"temporal","color":"#67b06e","indegree":5,"alsoIn":[],"x":0.1097,"y":0.3067,"text":"Active inference is a normative framework from computational neuroscience that unifies perception, action, and learning under a single imperative to maximize Bayesian model evidence, but prior implementations were restricted to low-dimensional discrete settings. The authors present a scalable implementation using amortized inference (neural networks) with Bayesian neural network transition models, cross-entropy method planning, and expected free energy decomposed into extrinsic value, state information gain, and parameter information gain. Results on continuous control benchmarks (MountainCar, inverted pendulum, hopper) demonstrate effective exploration and approximately an order of magnitude improvement in sample efficiency over DDPG. The work establishes formal homologies between active inference and state-of-the-art model-based RL."},{"id":"arxiv:2012.04293","kind":"paper","n":466,"label":"CRAFT: A Benchmark for Causal Reasoning About Forces and inTeractions","authors":"Ates et al.","year":2020,"url":"https://arxiv.org/abs/2012.04293","region":"t6","hemi":"llm","lobe":"temporal","color":"#b0a967","indegree":2,"alsoIn":[],"x":0.7123,"y":0.3149,"text":"CRAFT is a synthetic video question answering benchmark for causal and physical reasoning about forces and object interactions, built from ~57K question\u2013video pairs over 10K Box2D-simulated 2D videos across 20 scene layouts. Beyond previously-studied descriptive and counterfactual questions, it introduces a new causal question category grounded in the Force Dynamics Theory using cause/enable/prevent relations between affector and patient objects. The authors evaluate heuristic, text-only, single-frame, video (MAC, TVQA/TVQA+, G-SWM, LSTM-CNN) and oracle baselines against a human study. Results show a large gap (>29% in the hard setting) between humans and the best neural models, demonstrating that current models fail at the benchmark's causal/physical reasoning."},{"id":"arxiv:2406.07507","kind":"paper","n":467,"label":"Flow Map Matching with Stochastic Interpolants: A Mathematical Framework for Consistency Models","authors":"Nicholas M. Boffi, Michael S. Albergo, Eric Vanden-Eijnden","year":2024,"url":"https://arxiv.org/abs/2406.07507","region":"t3","hemi":"wm","lobe":"occipital","color":"#9a714f","indegree":7,"alsoIn":[],"x":0.3533,"y":0.1146,"text":"The paper addresses the computational cost of numerical integration in diffusion and flow-based generative models by introducing Flow Map Matching (FMM), a mathematical framework for learning the two-time flow map of a probability flow ODE. The framework provides both distillation-based objectives (Lagrangian and Eulerian) for learning from a pre-trained velocity field and a direct training objective that requires no pre-trained model, unifying consistency models, consistency trajectory models, and progressive distillation under one formalism. Experiments on CIFAR-10 and ImageNet-32 demonstrate that the approach achieves sample quality comparable to flow matching while reducing generation time by a factor of 10-20."},{"id":"arxiv:2603.20927","kind":"paper","n":468,"label":"Active Inference for Physical AI Agents -- An Engineering Perspective","authors":"(see preprint)","year":2026,"url":"https://arxiv.org/abs/2603.20927","region":"t14","hemi":"wm","lobe":"temporal","color":"#67b06e","indegree":3,"alsoIn":[],"x":0.1149,"y":0.3125,"text":"This paper argues that Active Inference (AIF), grounded in the Free Energy Principle, offers a principled unified framework for physical AI agents such as robots. It develops the argument through a chain from probability theory to Bayesian machine learning, variational inference, active inference, and finally reactive message passing on Forney-style factor graphs, showing that VFE minimization unifies perception, learning, planning, and control within a single computational objective. The paper further shows that coupled AIF agents can be coarse-grained into higher-level AIF agents, yielding computational homogeneity across scales. No benchmark comparisons are presented; the contribution is theoretical and architectural."},{"id":"arxiv:2508.08762","kind":"paper","n":469,"label":"Bio-Inspired Artificial Neural Networks based on Predictive Coding","authors":"","year":null,"url":"https://arxiv.org/abs/2508.08762","region":"t20","hemi":"llm","lobe":"frontal","color":"#67b0b0","indegree":7,"alsoIn":[],"x":0.6151,"y":0.6553,"text":"This tutorial-style Lecture Notes paper presents a self-contained derivation of Predictive Coding (PC) as a biologically plausible alternative to backpropagation (BP) for training artificial neural networks, where weight updates rely only on local pre-/post-synaptic information rather than a global error signal. Starting from information theory and variational inference, it derives PC from negative-free-energy (ELBO) optimization with a Gaussian generative model and a Dirac-delta variational posterior, yielding local update rules for value neurons, error neurons, and generative weights. It then shows the formal connections by which PC can approximate BP gradients (when the output covariance is fixed to identity and rescaled) and, with recurrent connections, approximates the Kalman Filter. A computational example on MNIST, FashionMNIST, and CIFAR10 shows PC is competitive with BP on small tasks but needs more FLOPs and longer training, and underperforms on larger/more complex settings."},{"id":"nature.com:c3aa5b0d9273","kind":"paper","n":470,"label":"Learning produces an orthogonalized state machine in the hippocampus","authors":"Weinan Sun, Johan Winnubst, Maanasa Natrajan, et al.","year":2025,"url":"https://www.nature.com/articles/s41586-024-08548-w","region":"t28","hemi":"llm","lobe":"frontal","color":"#7667b0","indegree":6,"alsoIn":[],"x":0.5568,"y":0.5151,"text":"This study investigates how cognitive maps form in the hippocampus during learning by recording thousands of CA1 neurons with two-photon calcium imaging while mice learned a virtual reality navigation task (2ACDC). The authors found that hippocampal activity undergoes progressive decorrelation of initially similar representations, ultimately forming an orthogonalized state machine (OSM) that captures latent task structure. Among computational models tested, the clone-structured causal graph (CSCG), an HMM variant, uniquely reproduced both the final orthogonalized representations and the step-by-step learning trajectory observed in animals, while LSTMs and transformers did not naturally produce such representations."},{"id":"arxiv:2511.15722","kind":"paper","n":471,"label":"Spatial Reasoning in Multimodal Large Language Models: A Survey of Tasks, Benchmarks and Methods","authors":"survey authors","year":2025,"url":"https://arxiv.org/abs/2511.15722","region":"t12","hemi":"llm","lobe":"frontal","color":"#5e9a4f","indegree":6,"alsoIn":[],"x":0.6166,"y":0.155,"text":"This survey introduces a cognitive-function-based taxonomy for spatial reasoning in MLLMs, organizing tasks along five cognitive categories (intrinsic/extrinsic, qualitative/quantitative, static/dynamic) and four levels of reasoning complexity (direct perception through advanced synthetic problems). The authors map 40+ existing benchmarks across text-only, vision-language, and embodied settings onto this taxonomy, review evaluation metrics, and analyze 40+ training- and inference-based methods. Key findings include a heavy concentration of benchmarks on extrinsic-qualitative-static reasoning, a critical gap in metric-level quantitative tasks, and the observation that chain-of-thought prompting can even hurt spatial reasoning performance on certain benchmarks."},{"id":"arxiv:2402.02716","kind":"paper","n":472,"label":"Understanding the planning of LLM agents: A survey","authors":"Xu Huang, Weiwen Liu, Xiaolong Chen, et al.","year":2024,"url":"https://arxiv.org/abs/2402.02716","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":9,"alsoIn":[],"x":0.5829,"y":0.9455,"text":"This survey provides the first systematic taxonomy of LLM-based agent planning, categorizing existing works into five directions: Task Decomposition, Multi-Plan Selection, External Planner-Aided Planning, Reflection and Refinement, and Memory-Augmented Planning. For each direction, the paper analyzes representative methods, their formulations, advantages, and limitations. The authors also conduct experiments on four benchmarks (ALFWorld, ScienceWorld, HotPotQA, FEVER) comparing six prompt-based methods, finding that performance scales with token expenses and that reflection mechanisms yield the largest gains on complex tasks."},{"id":"arxiv:2311.03658","kind":"paper","n":473,"label":"The Linear Representation Hypothesis and the Geometry of Large Language Models","authors":"Park, Choe & Veitch","year":2024,"url":"https://arxiv.org/abs/2311.03658","region":"t6","hemi":"llm","lobe":"temporal","color":"#b0a967","indegree":2,"alsoIn":[],"x":0.6884,"y":0.4714,"text":"The paper tackles the ambiguity of the 'linear representation hypothesis' in LLMs\u2014what 'linear representation' means and how to make geometric notions (cosine similarity, projection) meaningful in representation space. Using counterfactual pairs, the authors give two formalizations of linear representation (in the unembedding/output space and the embedding/input-context space) and prove the unembedding notion connects to linear probing (measurement) while the embedding notion connects to model steering (intervention). They introduce a 'causal inner product' that makes causally separable concepts orthogonal and unifies the two representation types via a Riesz isomorphism, with a closed-form estimator M = Cov(\u03b3)\u207b\u00b9. Experiments on LLaMA-2-7B over 27 concepts confirm concepts are linearly encoded as directions and that the causal inner product respects semantic structure."},{"id":"arxiv:1910.01442","kind":"paper","n":474,"label":"CLEVRER: CoLlision Events for Video REpresentation and Reasoning","authors":"Yi et al.","year":2020,"url":"https://arxiv.org/abs/1910.01442","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":1,"alsoIn":[],"x":0.3844,"y":0.6195,"text":"The paper introduces CLEVRER, a diagnostic synthetic video dataset of colliding objects (20,000 videos, >300,000 questions) designed to evaluate temporal and causal reasoning rather than perception, with four question types: descriptive, explanatory, predictive, and counterfactual. State-of-the-art visual reasoning models perform well on perception-based descriptive questions but poorly on the causal tasks (explanatory, predictive, counterfactual). The authors propose an oracle model, Neuro-Symbolic Dynamic Reasoning (NS-DR), that combines a neural video parser, a learned dynamics predictor (PropNet), a seq2seq question parser, and a symbolic program executor, which substantially outperforms baselines across all task types. The findings argue that causal video reasoning requires object-centric representations plus explicit dynamics/causal modeling, not just pattern recognition."},{"id":"arxiv:2606.09032","kind":"paper","n":475,"label":"Bridging the Agent-World Gap: Text World Models for LLM-based Agents","authors":"(authors listed on arXiv)","year":2026,"url":"https://arxiv.org/abs/2606.09032","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":8,"alsoIn":[],"x":0.6121,"y":0.9057,"text":"This survey provides the first systematic review of text world models (TWMs) for LLM-based agents\u2014transition models over textual states that, given a state and candidate action, predict the resulting webpage, terminal output, API response, or user reply. It organizes the field around a formal transition-function definition (M: S\u00d7A\u2192TS), a two-axis taxonomy (state representation \u00d7 grounding domain), and the agent lifecycle: construction (learning-based, prompt-based, programmatic/code-as-WM), training-time use (internalization, WM-as-environment, user simulation), inference-time use (simulator/lookahead, verifier), and evaluation (prediction fidelity, task utility, WM-as-eval-environment). It synthesizes scattered web/code/tool/dialogue work, surfacing trends (supervision moving from token fidelity to behavior-level signals; scaling shifting from trajectories to environments) and open problems like WM-policy coupling, reasoning world models, and grounding text predictions in physical reality."},{"id":"arxiv:2406.06385","kind":"paper","n":476,"label":"Low-Rank Quantization-Aware Training for LLMs (LR-QAT)","authors":"Yelysei Bondarenko, Riccardo Del Chiaro, Markus Nagel","year":2024,"url":"https://arxiv.org/abs/2406.06385","region":"t37","hemi":"llm","lobe":"parietal","color":"#b06788","indegree":3,"alsoIn":[],"x":0.8383,"y":0.3703,"text":"Quantization-aware training (QAT) yields better low-bit LLM quality than post-training quantization (PTQ) but is prohibitively memory- and compute-expensive. The paper proposes LR-QAT, a memory-efficient QAT method that freezes pretrained weights and trains low-rank adapters placed INSIDE the rounding/clipping operator so they align with the quantization grid and fuse into a single low-bit integer weight matrix at inference with zero overhead; it adds a fixed-point/double-packed INT8 downcasting operator and gradient checkpointing. LR-QAT matches full-model QAT accuracy at a fraction of the memory (reducing 7B LLaMA-2 QAT from 98.5GB to 20.5GB, trainable on a single 24GB consumer GPU) and outperforms PTQ baselines across LLaMA-1/2/3 and Mistral. It works across per-channel/group-wise weight quantization and weight-activation quantization, and can be combined with existing PTQ techniques."},{"id":"arxiv:2211.10438","kind":"paper","n":477,"label":"SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models","authors":"Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, Song Han","year":2022,"url":"https://arxiv.org/abs/2211.10438","region":"t37","hemi":"llm","lobe":"parietal","color":"#b06788","indegree":6,"alsoIn":[],"x":0.8706,"y":0.3937,"text":"Serving LLMs is memory- and compute-intensive, and INT8 W8A8 quantization is hard because large systematic activation outliers emerge above ~6.7B parameters, destroying accuracy under efficient per-tensor schemes. SmoothQuant is a training-free post-training quantization method that offline migrates quantization difficulty from activations to weights via a mathematically equivalent per-channel smoothing transformation (controlled by migration strength alpha), making both easy to quantize with hardware-friendly INT8 GEMM kernels. It preserves near-FP16 accuracy across OPT, BLOOM, GLM-130B, MT-NLG 530B, Llama-1/2, Falcon, Mistral, and Mixtral, delivering up to 1.56x speedup and ~2x memory reduction. It enables serving OPT-175B on half the GPUs and MT-NLG 530B within a single 8-GPU node."},{"id":"arxiv:2305.16291","kind":"paper","n":478,"label":"Voyager: An Open-Ended Embodied Agent with Large Language Models","authors":"Wang, Xie, Jiang, Mandlekar, Xiao, Zhu, Fan, Anandkumar","year":2023,"url":"https://arxiv.org/abs/2305.16291","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":0,"alsoIn":[],"x":0.7078,"y":0.7764,"text":"VOYAGER is the first LLM-powered embodied lifelong learning agent for Minecraft that explores, acquires skills, and makes discoveries without human intervention by querying GPT-4 via blackbox prompting (no fine-tuning). It combines three components: an automatic curriculum that maximizes exploration, an ever-growing skill library of executable code indexed by description embeddings, and an iterative prompting mechanism that refines generated programs using environment feedback, execution errors, and GPT-4 self-verification. Empirically it obtains 3.3\u00d7 more unique items, travels 2.3\u00d7 longer distances, and unlocks tech-tree milestones up to 15.3\u00d7 faster than prior SOTA, and is the only method to reach the diamond tier. It also generalizes its learned skill library zero-shot to solve novel tasks in a freshly instantiated world where baselines fail entirely."},{"id":"arxiv:2404.00610","kind":"paper","n":479,"label":"RQ-RAG: Learning to Refine Queries for Retrieval Augmented Generation","authors":"Chi-Min Chan et al.","year":2024,"url":"https://arxiv.org/abs/2404.00610","region":"t31","hemi":"llm","lobe":"occipital","color":"#9767b0","indegree":7,"alsoIn":[],"x":0.8798,"y":0.6356,"text":"LLMs with standard RAG retrieve context using the raw input query, which fails on ambiguous or complex queries and can hurt performance when irrelevant context is injected. This paper proposes RQ-RAG, which trains a 7B Llama2 model end-to-end on ~40k ChatGPT-annotated instances to explicitly rewrite, decompose, and disambiguate queries (via special/control tokens) and to regenerate answers grounded in retrieved contexts, plus three trajectory-selection strategies (PPL, confidence, ensemble). It surpasses the prior SOTA Self-RAG by 1.9% average across three single-hop QA datasets despite using far less training data, and improves substantially on multi-hop QA. Ablations show regenerating answers from retrieved context (0% original-answer retention) is best and the system is resilient across retrieval sources."},{"id":"arxiv:2105.04906","kind":"paper","n":480,"label":"VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning","authors":"Adrien Bardes, Jean Ponce, Yann LeCun","year":2021,"url":"https://arxiv.org/abs/2105.04906","region":"t23","hemi":"wm","lobe":"occipital","color":"#4f799a","indegree":0,"alsoIn":[],"x":0.2859,"y":0.2825,"text":"VICReg is a self-supervised joint-embedding method for image representation learning that prevents the collapse problem using three loss terms: an invariance term (MSE between two views' embeddings), a variance term (a hinge loss keeping each embedding dimension's standard deviation above a threshold), and a covariance term (decorrelating embedding dimensions by penalizing off-diagonal covariances). Unlike contrastive or distillation methods, it needs no negative pairs, memory bank, momentum encoder, stop-gradient, or batch/feature normalization, and applies its regularizers to each branch independently. It matches state-of-the-art on ImageNet linear evaluation (73.2% top-1, on par with Barlow Twins) and transfer tasks, while its branch-independent design enables multi-modal setups (image-text, audio) where it outperforms Barlow Twins. The variance term can also be added to BYOL/SimSiam to stabilize training and yield small gains."},{"id":"fi.ee.tsinghua.edu.cn:1c1fbee23e33","kind":"paper","n":481,"label":"A Survey of Embodied World Models","authors":"Shang et al.","year":2025,"url":"https://fi.ee.tsinghua.edu.cn/public/publications/0940dda4-af15-11f0-9d60-0242ac120002.pdf","region":"t17","hemi":"llm","lobe":"parietal","color":"#4f9a79","indegree":4,"alsoIn":[],"x":0.5955,"y":0.3759,"text":"This survey systematically reviews embodied world models\u2014systems that represent world states and model state transitions to help robots understand, interact with, and plan in physical environments. It proposes a technical taxonomy organizing the field along four axes: model architectures (video generation-based, 3D reconstruction-enhanced, and latent-space world models), training paradigms (instruction-conditioned, action-conditioned, physics-informed, video-action joint, and RL-based), application roles (offline data generation engine, RL environment substitute, robotic policy evaluator, and on-device action planner), and evaluation perspectives (generated data quality, end-to-end manipulation, policy-evaluation reliability, and downstream data scaling). It contrasts world models with VLA models\u2014arguing vision-centric world modeling is better suited to embodied learning long-term\u2014and outlines challenges in data collection, causal architecture, benchmarking, LLM integration, deployment, and building universal cross-scale physical world models."},{"id":"arxiv:2502.11946","kind":"paper","n":482,"label":"Step-Audio: Unified Understanding and Generation in Intelligent Speech Interaction","authors":"Ailin Huang et al.","year":2025,"url":"https://arxiv.org/abs/2502.11946","region":"t11","hemi":"llm","lobe":"occipital","color":"#81b067","indegree":3,"alsoIn":[],"x":0.8549,"y":0.3225,"text":"Step-Audio addresses limitations of open-source speech interaction systems\u2014high voice data collection costs, weak dynamic control, and limited intelligence\u2014by introducing a 130B-parameter unified speech-text multi-modal model with a dual-codebook tokenizer (linguistic + semantic), a generative speech data engine, and RLHF-based post-training. The model achieves SOTA on ASR, TTS, and audio question-answering benchmarks, with a 3B distilled TTS variant also released. A new benchmark, StepEval-Audio-360, is introduced for multi-dimensional evaluation of end-to-end speech dialogue."},{"id":"technologyreview.com:6e6b17c22c9b","kind":"paper","n":483,"label":"World models: 10 Things That Matter in AI Right Now","authors":"Will Douglas Heaven","year":2026,"url":"https://www.technologyreview.com/2026/04/21/1135650/world-models-ai-artificial-intelligence/","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":9,"alsoIn":[],"x":0.3118,"y":0.9114,"text":"This is a MIT Technology Review feature arguing that 'world models' \u2014 internal representations that let an intelligent system simulate its environment and predict the consequences of actions \u2014 are the key to extending AI from digital mastery to the physical world (robotics, navigation). It frames the current surge of interest around Google DeepMind, Fei-Fei Li's World Labs, Yann LeCun's new world-model startup after leaving Meta, and OpenAI reallocating Sora resources to 'world simulation' research. It contrasts world models with LLMs, citing a study where a model trained on simulated NYC taxi trips gives good navigation directions but fails completely when forced to take detours, suggesting LLM world understanding is brittle. It concludes that today's world models mostly generate interactive 3D environments for games/VR, but the real payoff would come from embedding them in agents that represent environments, predict action outcomes, and decide what to do."},{"id":"arxiv:2309.09777","kind":"paper","n":484,"label":"DriveDreamer: Towards Real-World-Driven World Models for Autonomous Driving","authors":"Wang et al.","year":2023,"url":"https://arxiv.org/abs/2309.09777","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":6,"alsoIn":[],"x":0.3948,"y":0.7937,"text":"DriveDreamer is presented as the first world model built entirely from real-world driving scenarios (nuScenes) rather than gaming or simulated environments. It introduces the Autonomous-driving Diffusion Model (Auto-DM), conditioned on structured traffic information (HDMaps, 3D boxes, text prompts, and driving actions), trained via a two-stage pipeline: first learning structured traffic constraints (image then video generation), then learning to predict future video and actions via an ActionFormer. The model achieves controllable driving video generation, synthetic-data augmentation that boosts 3D detection (e.g., BEVFusion mAP +3.0), and competitive open-loop planning (0.29m L2 average trajectory error)."},{"id":"arxiv:2506.22567","kind":"paper","n":485,"label":"Unifying Biomedical Vision-Language Expertise: Towards a Generalist Foundation Model via Multi-CLIP Knowledge Distillation","authors":"Shansong Wang et al.","year":2025,"url":"https://arxiv.org/abs/2506.22567","region":"t13","hemi":"llm","lobe":"parietal","color":"#6bb067","indegree":5,"alsoIn":[],"x":0.6675,"y":0.1052,"text":"Transferring CLIP's success to biomedicine is limited by scarce large-scale image-text corpora and heterogeneous imaging modalities, preventing a unified generalist model trained from scratch. The authors introduce MMKD-CLIP, which distills knowledge from nine state-of-the-art biomedical CLIP teacher models via a two-stage pipeline: contrastive pretraining on 2.9M image-text pairs (26 modalities) followed by feature-level distillation using 19.2M teacher feature quadruplets. Evaluated on 58 datasets (10.8M images, 9 modalities, 6 task types), MMKD-CLIP consistently outperforms all individual teacher models across zero-shot classification, linear probing, retrieval, VQA, survival prediction, and cancer diagnosis. The work shows multi-teacher knowledge distillation is a scalable paradigm for building high-performing biomedical foundation models under real-world data constraints."},{"id":"arxiv:2411.15594","kind":"paper","n":486,"label":"A Survey on LLM-as-a-Judge","authors":"Jiawei Gu et al.","year":2024,"url":"https://arxiv.org/abs/2411.15594","region":"t33","hemi":"llm","lobe":"parietal","color":"#ad67b0","indegree":4,"alsoIn":[],"x":0.8754,"y":0.6284,"text":"This survey systematizes the rapidly fragmented 'LLM-as-a-Judge' literature, providing formal and informal definitions, a contextualized definition of reliability (R \u2190 f_R(P_LLM, x, C)), and a four-question framework (what it is, how to use it, how to improve it, how to evaluate it). It catalogs improvement strategies across three pipeline phases (prompt/in-context design, model capability enhancement, output post-processing) and three evaluation dimensions (human agreement, bias, adversarial robustness). The authors run a meta-evaluation experiment on six LLMs and four common improvement strategies using LLMEval2 and EVALBIASBENCH, finding GPT-4 is the strongest, least-biased judge while many popular reliability strategies are ineffective. It closes with applications (finance, law, science, software, etc.), challenges, and a forward-looking research agenda including reasoning-centric and self-evolving judges."},{"id":"arxiv:1805.07526","kind":"paper","n":487,"label":"Deep Predictive Coding Network with Local Recurrent Processing for Object Recognition","authors":"Kuan Han, Haiguang Wen, Yizhen Zhang, Di Fu, Eugenio Culurciello, Zhongming Liu","year":2018,"url":"https://arxiv.org/abs/1805.07526","region":"t20","hemi":"llm","lobe":"frontal","color":"#67b0b0","indegree":1,"alsoIn":[],"x":0.5957,"y":0.5954,"text":"Inspired by the neuroscience theory of predictive coding, the paper proposes the Predictive Coding Network (PCN), a bi-directional CNN where feedback connections carry top-down predictions of lower-layer representations and feedforward connections carry bottom-up prediction errors, with adjacent layers interacting through local recurrent processing that iteratively refines representations via gradient descent on layer-wise squared prediction error. Unrolling this recurrent processing over T cycles turns a shallow static network into an effectively deep one without adding layers or parameters. On SVHN, CIFAR-10/100 and ImageNet, PCN matches or beats much deeper classical and SOTA models with far fewer layers and parameters (e.g., 9-layer PCN-D-5 reaches 21.77% error on CIFAR-100 vs 25.31% for its plain counterpart). Behavioral analysis shows internal representations converge over recurrent cycles, accuracy improves with more cycles, and the spatial pattern of prediction errors reveals visual saliency / bottom-up attention."},{"id":"arxiv:2402.12451","kind":"paper","n":488,"label":"The Revolution of Multimodal Large Language Models: A Survey","authors":"","year":null,"url":"https://arxiv.org/abs/2402.12451","region":"t30","hemi":"llm","lobe":"temporal","color":"#8c67b0","indegree":11,"alsoIn":[],"x":0.727,"y":0.2253,"text":"This survey reviews recent visual-based Multimodal Large Language Models (MLLMs), which connect frozen/fine-tuned visual encoders to LLM backbones via vision-to-language adapters and visual instruction tuning. It organizes the field along three axes \u2014 architecture (visual encoders, adapters such as linear/MLP projections, Q-Former, and gated cross-attention layers), training methodology (single- vs two-stage, alignment then instruction-tuning), and tasks (VQA, captioning, visual grounding, image generation/editing, plus video, 3D, and any-modality models). It additionally compiles training datasets and evaluation benchmarks and tabulates comparative performance and computational (GPU/TPU-hour) requirements across dozens of models."},{"id":"arxiv:2206.08853","kind":"paper","n":489,"label":"MineDojo: Building Open-Ended Embodied Agents with Internet-Scale Knowledge","authors":"Linxi Fan, Guanzhi Wang, Yunfan Jiang, Ajay Mandlekar, Yuncong Yang, Haoyi Zhu, ","year":2022,"url":"https://arxiv.org/abs/2206.08853","region":"t22","hemi":"llm","lobe":"temporal","color":"#4f849a","indegree":3,"alsoIn":[],"x":0.7049,"y":0.7585,"text":"Autonomous embodied agents typically train tabula rasa in narrow environments and fail to generalize. MineDojo introduces a framework built on Minecraft with a simulation suite of 1,000s of open-ended, language-prompted tasks plus an internet-scale multimodal knowledge base (730K+ YouTube videos with transcripts, 6K+ Wiki pages, 340K+ Reddit posts). The authors propose MineCLIP, a contrastive video-language model trained on 640K video-text pairs from the YouTube data, used as a learned dense, language-conditioned reward for RL. Agents trained with MineCLIP reward solve most of 12 tasks, match or beat hand-engineered dense rewards (up to 73% higher success in some cases), and MineCLIP doubles as an automatic evaluator that agrees well with human judgment."},{"id":"arxiv:2306.15668","kind":"paper","n":490,"label":"Physion++: Evaluating Physical Scene Understanding that Requires Online Inference of Different Physical Properties","authors":"Tung et al.","year":2023,"url":"https://arxiv.org/abs/2306.15668","region":"t6","hemi":"wm","lobe":"temporal","color":"#b0a967","indegree":2,"alsoIn":[],"x":0.4207,"y":0.815,"text":"Physion++ is a benchmark of 9 simulated video scenarios (built in the ThreeDWorld simulator) that test whether physical-prediction models can infer four latent mechanical properties \u2014 mass, friction, elasticity, and deformability \u2014 that are only observable through how objects move and interact, then use those inferences for an object-contact prediction (OCP) task. The authors evaluate seven state-of-the-art models spanning pixel-prediction, ImageNet-pretrained MLP readout, object-centric, and 3D particle-based architectures, and compare them against 200 human participants. They find models perform only slightly above chance and do not spontaneously learn latent-property inference, while humans reach ~60% accuracy; crucially, no model correlates well with human judgments, indicating models are not making predictions in a human-like way."},{"id":"arxiv:2403.08763","kind":"paper","n":491,"label":"Simple and Scalable Strategies to Continually Pre-train Large Language Models","authors":"Adam Ibrahim et al.","year":2024,"url":"https://arxiv.org/abs/2403.08763","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":8,"alsoIn":[],"x":0.7901,"y":0.6967,"text":"LLMs are typically re-trained from scratch on the union of old and new data whenever fresh data arrives, which is computationally wasteful, while naive continued training causes catastrophic forgetting or poor adaptation. This work shows that a simple combination of learning-rate re-warming, re-decaying, and replay of a small fraction of previous data lets continually pre-trained decoder-only transformers match the final validation loss and average benchmark score of full re-training, at a fraction of the compute. It validates this across 405M and 10B parameter models, weak (Pile\u2192SlimPajama, English\u2192English) and strong (Pile\u2192German) distribution shifts, and hundreds of billions of tokens. It also proposes infinite learning-rate schedules that avoid re-warming and are not tied to a fixed token budget."},{"id":"arxiv:2406.07472","kind":"paper","n":492,"label":"4Real: Towards Photorealistic 4D Scene Generation via Video Diffusion Models","authors":"Heng Yu, Chaoyang Wang, Peiye Zhuang, Willi Menapace, Aliaksandr Siarohin, et al","year":2024,"url":"https://arxiv.org/abs/2406.07472","region":"t15","hemi":"wm","lobe":"occipital","color":"#4f9a62","indegree":5,"alsoIn":[],"x":0.4492,"y":0.5643,"text":"4Real addresses photorealistic text-to-4D dynamic scene generation, avoiding the object-centric, non-photorealistic bias of prior methods that distill from multi-view generative models fine-tuned on synthetic object datasets. The method instead fully leverages a pre-trained text-to-video diffusion model (Snap Video), generating a reference video plus a camera-moving 'freeze-time' video, then reconstructing deformable 3D Gaussian Splats (D-3DGS) with a canonical representation, per-frame deformation to absorb multi-view inconsistencies, and temporal deformation for motion, refined via multi-view and temporal score distillation sampling. In user studies against 4Dfy and Dream-in-4D, 4Real wins on all seven criteria and improves X-CLIP and VideoScore metrics, while generating a scene in 1.5 hours on an A100 versus 10+ hours for competitors."},{"id":"arxiv:2410.00161","kind":"paper","n":493,"label":"KV-Compress: Paged KV-Cache Compression with Variable Compression Rates per Attention Head","authors":"Isaac Rehg","year":2024,"url":"https://arxiv.org/abs/2410.00161","region":"t18","hemi":"llm","lobe":"temporal","color":"#67b09a","indegree":5,"alsoIn":[],"x":0.8858,"y":0.4349,"text":"Long-context LLM inference is bottlenecked by KV cache memory, which scales with context length and limits concurrent request batching. KV-Compress modifies PagedAttention to evict contiguous KV blocks with variable compression rates per attention head and per layer, using squared attention metrics aggregated over query groups for GQA models, making variable-head-rate eviction practical in physical memory. The method achieves state-of-the-art on LongBench for Mistral-7B-Instruct-v0.2 and Llama-3.1-8B-Instruct while using 4x fewer KVs than baselines, and its vLLM integration yields up to 5.18x throughput improvement."},{"id":"arxiv:0910.3570","kind":"paper","n":494,"label":"The cascades route to chaos","authors":"","year":null,"url":"https://arxiv.org/abs/0910.3570","region":"t36","hemi":"llm","lobe":"frontal","color":"#b06793","indegree":5,"alsoIn":[],"x":0.663,"y":0.556,"text":"The paper develops a general theory explaining why period-doubling cascades (a classic route to chaos) appear, and proves that whenever a one- or two-dimensional smooth parametrized map transitions from no chaos to chaos as a parameter varies, there must be infinitely many cascades under mild genericity hypotheses (its 'chaos' includes non-attracting chaotic sets, defined as infinitely many regular periodic orbits). It distinguishes bounded cascades (which always come in pairs connected by an unstable periodic orbit) from unbounded cascades (carrying a constant-period 'stem'), and proves an Off-On-Off Chaos Theorem giving infinitely many bounded paired cascades. It establishes a conservation principle: unbounded cascades are invariant under large-scale perturbations, with exact stem-period counts computed for perturbed quadratic, cubic, and quartic maps. Numerical studies show the results apply to the forced-damped pendulum, double-well Duffing, Ikeda, and pulsed damped rotor maps."},{"id":"arxiv:2210.13382","kind":"paper","n":495,"label":"Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task","authors":"Li et al. (Kenneth Li, Aspen K. Hopkins, David Bau, Fernanda Vi\u00e9gas, Hanspeter P","year":2023,"url":"https://arxiv.org/abs/2210.13382","region":"t6","hemi":"wm","lobe":"temporal","color":"#b0a967","indegree":8,"alsoIn":[],"x":0.1667,"y":0.5115,"text":"The paper investigates whether sequence models build internal world representations or merely memorize surface statistics, using a GPT variant (Othello-GPT) trained only on Othello game-move transcripts to predict legal next moves. Probing experiments show that a nonlinear (2-layer MLP) probe recovers the board state from internal activations with high accuracy while linear probes fail, and interventional experiments confirm this representation is causally used by the model. The authors further use the intervention technique to build 'latent saliency maps' that explain predictions in terms of the board rather than input tokens. The finding is evidence for an emergent, nonlinear world model arising from pure next-token prediction with no a priori knowledge of game rules."},{"id":"arxiv:2308.01399","kind":"paper","n":496,"label":"Learning to Model the World with Language","authors":"Lin, Du, Watkins, Hafner, Abbeel, Klein & Dragan","year":2023,"url":"https://arxiv.org/abs/2308.01399","region":"t19","hemi":"wm","lobe":"occipital","color":"#67b0a5","indegree":6,"alsoIn":[],"x":0.203,"y":0.6889,"text":"Addresses how embodied agents can leverage diverse language (descriptions, game rules, corrections, general knowledge) beyond simple instructions and ground it to vision and action. Proposes Dynalang, which interprets language as a signal for predicting the future and learns a multimodal world model (built on DreamerV3) that compresses one image frame and one language token per timestep into a discrete latent, predicts future latent representations of text/image/reward, and trains an actor-critic purely from imagined rollouts. Across HomeGrid, Messenger, VLN-CE, and LangRoom, Dynalang outperforms model-free language-conditioned policies (IMPALA, R2D2, EMMA); crucially its performance improves with more diverse language while baselines degrade. The generative world model further enables text-only pretraining (e.g., on ~500M-token TinyStories) and embodied language generation."},{"id":"philpapers.org:f2d0d9f7c8fe","kind":"paper","n":497,"label":"Predictive Coding and Thought","authors":"Daniel Williams","year":2020,"url":"https://philpapers.org/rec/WILPCA-12","region":"t2","hemi":"llm","lobe":"temporal","color":"#b07d67","indegree":1,"alsoIn":[],"x":0.6121,"y":0.7317,"text":"The provided text contains no actual paper content \u2014 only a Cloudflare security block page from philpapers.org. Based solely on the title 'Predictive Coding and Thought,' the work appears to be a theoretical/philosophical treatment of predictive coding as an account of cognition and thought, but no method, experiments, or results can be extracted because the body text was not retrievable."},{"id":"arxiv:2509.20021","kind":"paper","n":498,"label":"Embodied AI: From LLMs to World Models","authors":"(multi-author)","year":2025,"url":"https://arxiv.org/abs/2509.20021","region":"t7","hemi":"wm","lobe":"occipital","color":"#adb067","indegree":7,"alsoIn":[],"x":0.2133,"y":0.8099,"text":"This invited survey reviews embodied AI spanning its history, key technologies (CV/NLP/RL/LLMs/WMs), core closed-loop components (active perception, embodied cognition, dynamic interaction), and hardware systems, then organizes the field into two burgeoning branches: LLM/MLLM-driven embodied AI (semantic reasoning + task decomposition) and World-Model-driven embodied AI (internal representations + future predictions). It argues each branch alone is insufficient \u2014 MLLMs ignore physical constraints while WMs lack high-level semantics \u2014 and proposes a joint MLLM-WM-driven embodied AI architecture that couples semantic planning with physics-grounded simulation and memory updating. It supports the argument with taxonomic comparison tables of perception/cognition/interaction methods and a qualitative MLLM-only vs WM-only vs joint comparison, citing EvoAgent as an exemplar of the joint paradigm. The paper closes with representative applications (service robots, rescue UAVs, industrial robots) and future directions including autonomous, hardware, swarm, and trustworthy embodied AI."},{"id":"arxiv:1711.00937","kind":"paper","n":499,"label":"Neural Discrete Representation Learning","authors":"Aaron van den Oord, Oriol Vinyals, Koray Kavukcuoglu","year":2017,"url":"https://arxiv.org/abs/1711.00937","region":"t20","hemi":"wm","lobe":"frontal","color":"#67b0b0","indegree":2,"alsoIn":[],"x":0.3248,"y":0.2953,"text":"The paper introduces VQ-VAE, a variational autoencoder whose encoder emits discrete codes via nearest-neighbour lookup into a learned embedding table (vector quantisation), trained with a straight-through gradient estimator plus codebook and commitment loss terms. This discretisation sidesteps 'posterior collapse' when paired with powerful autoregressive decoders/priors (PixelCNN for images, WaveNet for audio). VQ-VAE matches continuous-latent VAEs on CIFAR10 log-likelihood (4.67 vs 4.51 bits/dim) and, when paired with an autoregressive prior, generates coherent 128x128 images, action-conditional video, and speech, while learning phoneme-like units and enabling unsupervised speaker conversion."},{"id":"arxiv:2410.08893","kind":"paper","n":500,"label":"Drama: Mamba-Enabled Model-Based Reinforcement Learning Is Sample and Parameter Efficient","authors":"Wang, Wang, van der Schaar et al.","year":2025,"url":"https://arxiv.org/abs/2410.08893","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":0,"alsoIn":[],"x":0.1304,"y":0.3412,"text":"The paper introduces Drama, the first model-based reinforcement learning agent that uses the Mamba (specifically Mamba-2) state space model as its world-model sequence backbone, achieving O(n) memory and compute scaling versus the O(n^2) cost of transformer attention while capturing long-term dependencies. It also proposes Dynamic Frequency-based Sampling (DFS), which preferentially samples transitions the world model has already learned to reduce suboptimality from inaccurate early-stage models. With only a 7M-parameter world model, Drama reaches a normalised mean score competitive with SOTA MBRL methods on the Atari100k benchmark and is trainable on a standard laptop. Ablations show Mamba-2 outperforms Mamba, DFS beats uniform sampling, and Mamba variants beat GRU/Transformer on a long-sequence grid-world task."},{"id":"arxiv:cs/0405001","kind":"paper","n":501,"label":"Toward a New Policy for Scientific and Technical Communication: the Case of Kyrgyz Republic","authors":"","year":null,"url":"https://arxiv.org/abs/cs/0405001","region":"t10","hemi":"llm","lobe":"temporal","color":"#8cb067","indegree":0,"alsoIn":[],"x":0.7063,"y":0.7218,"text":"This policy paper analyzes the state of scientific and technical information (STI) communication in the Kyrgyz Republic amid the global shift from paper-based to electronic scholarly communication, and proposes a new national STI policy. Drawing on a literature review plus interviews, a quantitative survey, and focus groups with over 100 scientists in Bishkek, Osh, and Kara-Kol, it identifies the core problems as inadequate access to information for Kyrgyz scientists and poor global visibility (\"lost science\") of locally produced research. It weighs two policy dimensions \u2014 electronic vs. print communication and a national STI system vs. cross-national virtual collaboration (collaboratories) \u2014 and recommends a balanced, predominantly electronic approach. The paper concludes with concrete recommendations across five policy components (telecommunications, computerization, STI systems, legislation, education) mapped to specific government, academic, library, publisher, and donor stakeholders."},{"id":"pi.website:ca5c1a691358","kind":"paper","n":502,"label":"A Cambrian Explosion of Robotics Applications","authors":"Physical Intelligence","year":2026,"url":"https://www.pi.website/blog","region":"t0","hemi":"wm","lobe":"frontal","color":"#9a4f4f","indegree":5,"alsoIn":[],"x":0.1146,"y":0.7537,"text":"This is a chronological index of Physical Intelligence's research releases (Oct 2024\u2013Apr 2026) on vision-language-action (VLA) robotic foundation models, tracing the evolution from the first generalist policy \u03c00 through \u03c00.5, \u03c00.6, and \u03c00.7. The work spans efficient action tokenization (FAST), real-time action chunking under latency, online RL for learning from experience, multi-scale embodied memory, human-to-robot transfer, and steerable models with emergent generalization. The unifying thesis is that general-purpose 'physical intelligence' foundation models will trigger a 'Cambrian explosion' of real-world robotics applications. Reported advances include 5x faster training via FAST tokenization and open-world generalization that lets a mobile manipulator clean an entirely new kitchen or bedroom."},{"id":"arxiv:1909.10863","kind":"paper","n":503,"label":"Active inference: demystified and compared","authors":"","year":null,"url":"https://arxiv.org/abs/1909.10863","region":"t14","hemi":"llm","lobe":"temporal","color":"#67b06e","indegree":6,"alsoIn":[],"x":0.5568,"y":0.9061,"text":"This paper provides an accessible tutorial of the discrete-state, discrete-time formulation of active inference and contrasts it directly with reinforcement learning on a modified 3x3 OpenAI gym FrozenLake task. It derives variational and expected free energy, shows that exploration/exploitation, epistemic foraging, and uncertainty handling emerge naturally rather than being hand-engineered, and that reward becomes just another preferred observation (prior preference) that agents can even learn. Empirically it compares active inference against Q-learning (fixed and decaying epsilon) and Bayesian model-based RL (Thompson sampling) in stationary, non-stationary, and reward-free settings. The key finding is that active inference matches RL when rewards are well-specified but markedly outperforms it under non-stationarity and continues to produce purposeful information-seeking behavior when rewards are absent."},{"id":"arxiv:2304.07193","kind":"paper","n":504,"label":"DINOv2: Learning Robust Visual Features without Supervision","authors":"Maxime Oquab, Timoth\u00e9e Darcet, Th\u00e9o Moutakanni, Huy V. Vo, Marc Szafraniec, Vasi","year":2023,"url":"https://arxiv.org/abs/2304.07193","region":"t23","hemi":"wm","lobe":"occipital","color":"#4f799a","indegree":1,"alsoIn":[],"x":0.2407,"y":0.3462,"text":"DINOv2 shows that discriminative self-supervised learning can produce general-purpose visual features that rival weakly-supervised models, without finetuning, when trained on enough curated data. The authors build an automatic curation pipeline to assemble LVD-142M (142M images) from uncurated web data via self-supervised retrieval, combine DINO/iBOT losses with SwAV centering, KoLeo regularization, and high-resolution adaptation, and add efficiency improvements (custom FlashAttention, sequence packing, FSDP) that make training ~2x faster with 1/3 the memory. They train a 1.1B-parameter ViT-g/14 and distill it into smaller ViTs. Frozen DINOv2 features surpass the best self-supervised models by large margins and match or exceed OpenCLIP across classification, segmentation, depth, retrieval and video tasks."},{"id":"arxiv:2510.17111","kind":"paper","n":505,"label":"Efficient Vision-Language-Action Models for Embodied Manipulation: A Systematic Survey","authors":"et al.","year":2025,"url":"https://arxiv.org/abs/2510.17111","region":"t0","hemi":"wm","lobe":"frontal","color":"#9a4f4f","indegree":11,"alsoIn":[],"x":0.2263,"y":0.8339,"text":"This survey is the first dedicated review of efficiency optimization for Vision-Language-Action (VLA) models, which map language instructions and visual observations to robot actions but suffer from large parameter counts, high memory use, and slow inference that conflict with edge/real-time deployment. It organizes existing efficiency techniques into four dimensions\u2014model architecture, perception feature, action generation, and training/inference strategies\u2014and summarizes representative methods within each. It analyzes the strengths and weaknesses of each strategy and outlines five future directions (model-data co-optimization, efficient 3D spatio-temporal perception, compact continuous action representations, imitation-to-RL adaptation, and efficiency-centric benchmarking)."},{"id":"arxiv:2603.21354","kind":"paper","n":506,"label":"The Workload-Router-Pool Architecture for LLM Inference Optimization: A Vision Paper from the vLLM Semantic Router Project","authors":"","year":null,"url":"https://arxiv.org/abs/2603.21354","region":"t39","hemi":"wm","lobe":"occipital","color":"#9a4f5b","indegree":1,"alsoIn":[],"x":0.1952,"y":0.4248,"text":"This paper argues that LLM inference optimization along three dimensions\u2014Workload (what the fleet serves), Router (how requests are dispatched), and Pool (where inference runs)\u2014are coupled rather than orthogonal, and that isolated per-dimension optimization leaves significant efficiency unrealized. It distills the vLLM Semantic Router project's prior publications into the Workload\u2013Router\u2013Pool (WRP) architecture, maps them onto a 3\u00d73 interaction matrix to identify covered and open cells, and proposes 21 concrete research directions at the intersections, tiered from engineering-ready to open research. The framework is grounded in the project's own measurements across routing mechanisms, fleet provisioning, multimodal/agent routing, and governance standards."},{"id":"arxiv:2511.02097","kind":"paper","n":507,"label":"A Step Toward World Models: A Survey on Robotic Manipulation","authors":"Yuxiang Ma, Peng-Fei Zhang, et al.","year":2025,"url":"https://arxiv.org/abs/2511.02097","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":7,"alsoIn":[],"x":0.3076,"y":0.842,"text":"This survey examines world models through the lens of robotic manipulation rather than committing to a single fixed definition, arguing that LLMs, VLMs, VLAs, and video generation models all exhibit core world-modeling characteristics. It organizes existing methods into three paradigms (implicit world modeling, latent dynamics modeling, and video generation), two architectural designs (flat vs. hierarchical), and analyzes world observation/representation and task scope, plus two functional roles: decision support (action prediction and planning) and training facilitation (data engine and evaluation). It distills key techniques and challenges\u2014data limitations, perception, long-horizon reasoning, spatiotemporal consistency, generalization, physics-informed learning, and memory\u2014and proposes 13 core components/capabilities a fully realized world model should possess. It compiles representative models (Table I, ~60 systems) and datasets (Table II) and outlines future directions including multimodal perception, hierarchical models, causality, lightweight deployment, safety, and unified evaluation."},{"id":"arxiv:2510.25760","kind":"paper","n":508,"label":"Multimodal Spatial Reasoning in the Large Model Era: A Survey and Benchmarks","authors":"Anonymous et al.","year":2025,"url":"https://arxiv.org/abs/2510.25760","region":"t12","hemi":"llm","lobe":"frontal","color":"#5e9a4f","indegree":7,"alsoIn":[],"x":0.6226,"y":0.147,"text":"This survey provides a comprehensive review of multimodal spatial reasoning with large models, covering test-time scaling, post-training (SFT/RL), architectural modifications, and explainability for MLLMs, then extending to 3D grounding/QA/generation, embodied AI (VLA, VLN, EQA, grasping, world models), and novel modalities (video, audio). It introduces a taxonomy unifying these areas and provides open benchmarks for evaluation. The authors also benchmark several VLM backbones used in VLA models on spatial reasoning tasks, finding that these backbones exhibit limited but non-trivial spatial reasoning abilities."},{"id":"arxiv:2505.14146","kind":"paper","n":509,"label":"s3: You Don't Need That Much Data to Train a Search Agent via RL","authors":"Pengcheng Jiang et al.","year":2025,"url":"https://arxiv.org/abs/2505.14146","region":"t31","hemi":"llm","lobe":"occipital","color":"#9767b0","indegree":4,"alsoIn":[],"x":0.8499,"y":0.7163,"text":"RAG search agents trained end-to-end with RL entangle retrieval with generation and require full model access, while retrieval-only RL optimizes search metrics (NDCG, recall) disconnected from downstream answer quality. This paper proposes s3, a model-agnostic framework that freezes the generator and trains only a search agent (Qwen2.5-7B) with PPO using a 'Gain Beyond RAG' (GBR) reward \u2014 the improvement in generation accuracy over na\u00efve top-k RAG. Trained on just 2.4k examples (20 PPO steps), s3 outperforms DeepRetrieval (70k) and Search-R1 (170k) \u2014 over 70\u00d7 more data \u2014 across six general QA and five medical QA benchmarks. The decoupled design transfers zero-shot to the medical domain and reduces training wall-clock by ~33\u00d7."},{"id":"arxiv:2303.07109","kind":"paper","n":510,"label":"Transformer-based World Models Are Happy With 100k Interactions (TWM)","authors":"Robine et al.","year":2023,"url":"https://arxiv.org/abs/2303.07109","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":5,"alsoIn":[],"x":0.1252,"y":0.4189,"text":"Deep RL agents are far more data-hungry than humans, so the paper builds a sample-efficient world model by applying a Transformer-XL autoregressively over real-environment episodes, feeding in compact discrete latent states, actions, AND experienced/predicted rewards as separate modalities. The transformer computes deterministic hidden states that predict reward, discount, and next latent state, generating imagined trajectories used to train a model-free actor-critic policy in latent space. Because the policy is conditioned only on the latent state z (with frame stacking), the transformer is not needed at inference, keeping the deployed policy cheap. On the Atari 100k benchmark TWM outperforms prior model-free and model-based methods, achieving a normalized mean of 0.956 and median of 0.505."},{"id":"arxiv:2506.01622","kind":"paper","n":511,"label":"General agents need world models","authors":"Jonathan Richens, Tom Everitt, David Abel","year":2025,"url":"https://arxiv.org/abs/2506.01622","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":1,"alsoIn":[],"x":0.7258,"y":0.7461,"text":"The paper asks whether world models are necessary for general, goal-directed agents or whether model-free learning suffices. It proves formally that any agent satisfying a regret bound over a sufficiently diverse set of multi-step goal-directed tasks must have learned an accurate predictive model of its environment, and derives an algorithm (Algorithm 1/2) that extracts this transition function (world model) from the agent's policy alone. The recovery error shrinks as agent regret \u03b4 decreases or maximum goal depth n increases, scaling as O(\u03b4/\u221an)+O(1/n); myopic (depth-1) agents are proven to need no world model. Experiments in a 20-state, 5-action cMP confirm the recovered model's average error decays as O(n^-1/2) even when the agent violates the regret-bound assumptions."},{"id":"arxiv:2303.03378","kind":"paper","n":512,"label":"PaLM-E: An Embodied Multimodal Language Model","authors":"Driess, Xia, Sajjadi, Lynch et al.","year":2023,"url":"https://arxiv.org/abs/2303.03378","region":"t0","hemi":"llm","lobe":"frontal","color":"#9a4f4f","indegree":0,"alsoIn":[],"x":0.6495,"y":0.2293,"text":"LLMs reason well over text but lack grounding in real-world sensor data, which limits their use for robotics. PaLM-E injects continuous observations (images, state estimates, and 3D neural scene representations) directly into the embedding space of a pre-trained PaLM LLM, forming 'multimodal sentences' trained end-to-end for embodied planning, VQA, and captioning across three robot embodiments. A single model exhibits positive transfer\u2014co-training on diverse internet-scale vision-language plus robot data more than doubles low-data robot planning performance versus in-domain-only training. The largest variant, PaLM-E-562B (540B PaLM + 22B ViT), sets state-of-the-art on OK-VQA and shows emergent zero-shot multimodal chain-of-thought and multi-image reasoning despite single-image training."},{"id":"arxiv:2406.15836","kind":"paper","n":513,"label":"Decentralized Transformers with Centralized Aggregation are Sample-Efficient Multi-Agent World Models","authors":"Yang Zhang, Chenjia Bai, Bin Zhao, Junchi Yan, Xiu Li, Xuelong Li","year":2024,"url":"https://arxiv.org/abs/2406.15836","region":"t22","hemi":"llm","lobe":"temporal","color":"#4f849a","indegree":9,"alsoIn":[],"x":0.6734,"y":0.7229,"text":"The paper tackles sample-inefficient multi-agent RL by building MARIE, the first Transformer-based multi-agent world model that learns decentralized local dynamics (for scalability across agent counts) combined with a centralized Perceiver-based representation aggregation across agents (to handle non-stationarity), mirroring the CTDE principle. Local observations are discretized with a shared VQ-VAE into tokens, and a shared GPT-style Transformer autoregressively predicts each agent's next observation tokens, reward, and discount, while a Perceiver compresses the joint n(K+1)-length token sequence into n per-agent global features. Policies are trained purely in imagination via a MAPPO-like actor-critic, decoupled from the world model for fast deployment. On SMAC in a low-data regime and on MAMuJoCo, MARIE outperforms model-free (MAPPO, QMIX, QPLEX) and model-based (MAMBA, MBVD) baselines, with gains growing on harder scenarios."},{"id":"arxiv:2305.14909","kind":"paper","n":514,"label":"Leveraging Pre-trained Large Language Models to Construct and Utilize World Models for Model-based Task Planning","authors":"Lin Guan, Karthik Valmeekam, Sarath Sreedharan, Subbarao Kambhampati","year":2023,"url":"https://arxiv.org/abs/2305.14909","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":6,"alsoIn":[],"x":0.7662,"y":0.8238,"text":"The paper tackles the unreliability of using LLMs directly as planners (incorrect plans, heavy reliance on execution feedback, repeated errors) by instead using LLMs to construct an explicit symbolic world model in PDDL, which is then handed to sound domain-independent planners. LLMs generate PDDL action models action-by-action with a shared predicate list, and serve as a natural-language interface between the PDDL and corrective feedback sources (a PDDL/VAL validator and human domain experts) so non-experts can fix the model up front. On two IPC domains (Logistics, Tyreworld) and a custom 22-skill Household domain, GPT-4 produces high-quality PDDL for 41 actions (400+ literals) with relatively few errors, and the corrected models let a classical planner solve 48 tasks. Fast Downward with the LLM-acquired PDDL reaches 95% (Household) and 100% (Logistics) success, vastly outperforming an LLM-only planner (15% / 0%)."},{"id":"arxiv:2406.06485","kind":"paper","n":515,"label":"Can Language Models Serve as Text-Based World Simulators?","authors":"Wang et al.","year":2024,"url":"https://arxiv.org/abs/2406.06485","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":3,"alsoIn":[],"x":0.648,"y":0.8876,"text":"The paper asks whether LLMs can serve directly as text-based world simulators, predicting how actions and environment dynamics change world states. The authors build BYTESIZED32-State-Prediction (BYTESIZED32-SP), a benchmark of 76,369 text-game state transitions from 31 games, and define the LLM-as-a-Simulator (LLM-Sim) task that decomposes the transition function into action-driven, environment-driven, and game-progress prediction over JSON states. Testing GPT-4 (and GPT-3.5), they find accuracy never exceeds 59.9% on transitions involving non-trivial state changes, with environment-driven transitions and arithmetic/common-sense/scientific reasoning being the main failure points. They conclude LLMs are not yet reliable world simulators, since errors compound across steps (~0.599^10 < 1% after 10 steps)."},{"id":"arxiv:2504.15785","kind":"paper","n":516,"label":"WALL-E 2.0: World Alignment by NeuroSymbolic Learning improves World Model-based LLM Agents","authors":"Zhou, Zhou, Yang et al.","year":2025,"url":"https://arxiv.org/abs/2504.15785","region":"t19","hemi":"wm","lobe":"occipital","color":"#67b0a5","indegree":1,"alsoIn":[],"x":0.215,"y":0.5945,"text":"LLM agents fail in open-world environments because of the gap between an LLM's prior knowledge and the specific environment's dynamics. The paper proposes WALL-E 2.0, a training-free 'world alignment' method that uses LLM inductive reasoning to extract environment-specific symbolic knowledge (action rules, knowledge graphs, scene graphs) from exploration trajectories, encodes it as executable Python code rules (pruned via maximum set coverage), and combines it with the LLM to form a neurosymbolic world model. An RL-free model-based agent then uses this world model inside an LLM-based model-predictive control (MPC) loop where the LLM acts as a look-ahead optimizer. On Mars (Minecraft-like) and ALFWorld, WALL-E 2.0 beats baselines by 16.1%\u201351.6% in success/reward and reaches a record 98% success on ALFWorld after only 4 iterations."},{"id":"arxiv:2309.07870","kind":"paper","n":517,"label":"Agents: An Open-source Framework for Autonomous Language Agents","authors":"","year":null,"url":"https://arxiv.org/abs/2309.07870","region":"t35","hemi":"llm","lobe":"occipital","color":"#b0679e","indegree":3,"alsoIn":[],"x":0.645,"y":0.9286,"text":"The paper introduces AGENTS, an open-source library/framework for building autonomous LLM-powered language agents that is accessible to non-specialists while remaining extensible for researchers. The framework is built around three core classes (Agent, Environment, and SOP) all configured from plain-text config files, and supports planning, long-short term memory, tool usage and web navigation, multi-agent communication, human-agent interaction, and fine-grained symbolic control. Its key novelties are symbolic plans called SOPs (standard operating procedures) \u2014 state graphs with LLM-driven transitions for controllable agent behavior \u2014 plus dynamic multi-agent scheduling, an Agent Hub for sharing agents, FastAPI deployment, and an automatic RAG-based SOP-generation 'meta agent'. The paper validates the framework through qualitative case studies (single-agent customer service/sales bots and multi-agent fiction studio, debate, and software-company systems) rather than quantitative benchmarks."},{"id":"arxiv:2201.11903","kind":"paper","n":518,"label":"Chain-of-Thought Prompting Elicits Reasoning in Large Language Models","authors":"Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, E","year":2022,"url":"https://arxiv.org/abs/2201.11903","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":3,"alsoIn":[],"x":0.835,"y":0.8241,"text":"The paper introduces chain-of-thought (CoT) prompting, in which a few few-shot exemplars are augmented with intermediate natural-language reasoning steps before the final answer, eliciting multi-step reasoning from off-the-shelf large language models without any finetuning. Across arithmetic, commonsense, and symbolic reasoning benchmarks on GPT-3, LaMDA, PaLM, UL2, and Codex, CoT prompting substantially outperforms standard prompting. The gains are an emergent ability of scale, appearing only at ~100B+ parameters. Most strikingly, PaLM 540B with eight CoT exemplars reaches new state-of-the-art on GSM8K, surpassing a finetuned GPT-3 with a verifier."},{"id":"arxiv:2603.03251","kind":"paper","n":519,"label":"Speculative Speculative Decoding","authors":"","year":null,"url":"https://arxiv.org/abs/2603.03251","region":"t9","hemi":"llm","lobe":"parietal","color":"#97b067","indegree":5,"alsoIn":[],"x":0.8683,"y":0.7762,"text":"Standard speculative decoding (SD) accelerates LLM inference but retains a sequential dependence: verification must finish before the next draft begins. This paper introduces speculative speculative decoding (SSD), which runs the draft model on separate hardware and, while verification is ongoing, predicts likely verification outcomes and pre-speculates for all of them in parallel, returning cached speculations instantly on a hit and falling back to just-in-time drafting on a miss. The authors identify three challenges (predicting the bonus token, the cache-hit vs acceptance-rate tradeoff, and miss handling) and solve each in an optimized lossless algorithm called Saguaro. Saguaro is on average ~30% faster than the strongest SD baselines and up to ~5x faster than autoregressive decoding across math, code, and chat datasets and two model families."},{"id":"arxiv:2507.05169","kind":"paper","n":520,"label":"Critique of World Model","authors":"Eric Xing, Mingkai Deng, Jinyu Hou, Zhiting Hu","year":2025,"url":"https://arxiv.org/abs/2507.05169","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":2,"alsoIn":[],"x":0.2859,"y":0.8993,"text":"A position/critique essay arguing that a world model is fundamentally a simulator of all actionable possibilities of the real world for purposeful 'simulative reasoning' (hypothetical thinking), NOT a video/content generator. It surveys the current landscape (gaming, 3D-scene, physical, video-generation, and JEPA-style world models) along five design dimensions \u2014 data, representation, architecture, objective, and usage \u2014 and critiques the dominant 'common wisdom' (sensory-first, continuous embeddings, encoder-encoder latent prediction, latent reconstruction loss, MPC). It proposes the Generative Latent Prediction (GLP) architecture as a synthesis: a stateful, hierarchical, mixed continuous/discrete representation with an enhanced-LLM backbone plus a diffusion embedding predictor, trained with an observation-grounded generative loss and used to simulate experience for RL. The argument is backed by formal results (Theorem 1 on completeness of discrete representations; Propositions 1\u20132 on collapse of latent loss vs non-collapse of generative loss; Theorem 2 showing latent reconstruction is an upper-bounded surrogate of generative reconstruction) and previews the PAN (Physical, Agentic, Nested) world model."},{"id":"arxiv:2307.06281","kind":"paper","n":521,"label":"MMBench: Is Your Multi-modal Model an All-around Player?","authors":"Yuan Liu et al.","year":2023,"url":"https://arxiv.org/abs/2307.06281","region":"t30","hemi":"llm","lobe":"temporal","color":"#8c67b0","indegree":2,"alsoIn":[],"x":0.8369,"y":0.2842,"text":"Evaluating large vision-language models (VLMs) is hard: exact-match public benchmarks (VQAv2, COCO Caption) cause false negatives and lack fine-grained ability analysis, while human subjective benchmarks (OwlEval) are biased and unscalable. The paper proposes MMBench, a bilingual (English/Chinese) objective benchmark of 3,217 multiple-choice questions spanning a hierarchical taxonomy of 20 fine-grained ability dimensions (2 L-1, 6 L-2, 20 L-3), built with LLM/VLM-based quality control. It introduces CircularEval (each question tested N times with circularly shifted choices, requiring all passes correct) and uses GPT-4 to extract choices from free-form predictions. Evaluating 21+ VLMs, InternLM-XComposer2 leads open-source models at 78.1% overall on MMBench-test, ahead of proprietary GPT-4v (74.3%) and Qwen-VL-Max (75.4%)."},{"id":"arxiv:2412.14161","kind":"paper","n":522,"label":"GAIA companion: SmolAgents and the WebSailor lineage aside \u2014 OSWorld-Verified omitted; included: TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks","authors":"Frank F. Xu et al.","year":2024,"url":"https://arxiv.org/abs/2412.14161","region":"t35","hemi":"llm","lobe":"occipital","color":"#b0679e","indegree":6,"alsoIn":[],"x":0.6246,"y":0.9586,"text":"The paper introduces TheAgentCompany, an extensible, self-hosted benchmark of 175 diverse professional tasks simulating a small software company, where LLM agents must browse the web, write and run code, and communicate with LLM-simulated colleagues across GitLab, OwnCloud, Plane, and RocketChat. Tasks use checkpoint-based evaluators (deterministic Python plus LLM-as-judge fallback) that award partial credit. Testing twelve model backbones via the OpenHands CodeAct and OWL-RolePlay harnesses, the best model (Gemini-2.5-Pro) autonomously completes only 30.3% of tasks (39.3% with partial credit), revealing that simpler tasks are automatable but long-horizon, social, and complex-UI tasks remain out of reach."},{"id":"arxiv:2307.12698","kind":"paper","n":523,"label":"MC-JEPA: A Joint-Embedding Predictive Architecture for Self-Supervised Learning of Motion and Content Features","authors":"Adrien Bardes, Jean Ponce, Yann LeCun","year":2023,"url":"https://arxiv.org/abs/2307.12698","region":"t23","hemi":"wm","lobe":"occipital","color":"#4f799a","indegree":4,"alsoIn":[],"x":0.273,"y":0.4212,"text":"Self-supervised visual representation learning typically captures content features (object identity) but ignores pixel-level motion, while optical flow estimation captures motion without understanding content. This paper introduces MC-JEPA, a joint-embedding predictive architecture that jointly learns self-supervised optical flow (M-JEPA, based on PWC-Net with added backward cycle-consistency and variance-covariance regularization) and content features (via VICReg on ImageNet) within a single shared ConvNeXt-T encoder in a multi-task setup. The two objectives benefit each other: the flow pretext task improves localization for segmentation, and content learning improves the estimated flow. MC-JEPA achieves optical flow on par with specialized unsupervised methods (e.g. KITTI 2015 train EPE 2.67) while matching or beating dedicated SSL methods on image and video segmentation with one encoder."},{"id":"arxiv:1811.04551","kind":"paper","n":524,"label":"Learning Latent Dynamics for Planning from Pixels (PlaNet)","authors":"Hafner et al.","year":2019,"url":"https://arxiv.org/abs/1811.04551","region":"t17","hemi":"wm","lobe":"parietal","color":"#4f9a79","indegree":2,"alsoIn":[],"x":0.1356,"y":0.4002,"text":"PlaNet (Deep Planning Network) is a purely model-based RL agent that learns environment dynamics from raw 64\u00d764\u00d73 pixel observations and selects actions via fast online planning (model-predictive control with the cross-entropy method) entirely in a compact latent space, with no policy or value network. The dynamics are modeled by a Recurrent State Space Model (RSSM) that combines deterministic (GRU) and stochastic transition paths, and the paper introduces 'latent overshooting,' a multi-step variational objective that trains predictions of all distances purely in latent space. On six DeepMind Control Suite tasks, PlaNet matches or beats strong model-free agents (A3C, D4PG) while using roughly 200\u00d7 fewer episodes (1,000 vs 100,000). Ablations show both stochastic and deterministic transition components are essential, and online data collection plus iterative CEM planning are needed for top performance."},{"id":"arxiv:2103.00020","kind":"paper","n":525,"label":"Learning Transferable Visual Models From Natural Language Supervision","authors":"Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini","year":2021,"url":"https://arxiv.org/abs/2103.00020","region":"t23","hemi":"wm","lobe":"occipital","color":"#4f799a","indegree":6,"alsoIn":[],"x":0.3411,"y":0.3619,"text":"CLIP tackles the limited generality of vision systems trained on fixed label sets by learning directly from natural language supervision. The method pre-trains an image encoder and a text encoder jointly with a contrastive objective\u2014predicting which caption goes with which image\u2014over 400M (image, text) pairs scraped from the web (the WIT dataset), then performs zero-shot classification by synthesizing a linear classifier from class-name text. Zero-shot CLIP matches the original ResNet-50's ImageNet accuracy (76.2% top-1) without using any of its 1.28M labeled examples, is competitive with fully supervised baselines across 30+ datasets, and is substantially more robust to natural distribution shift. The paper also studies scaling, prompt engineering, linear-probe representation quality, and social biases."},{"id":"arxiv:2506.18701","kind":"paper","n":526,"label":"Matrix-Game: Interactive World Foundation Model","authors":"Skywork AI","year":2025,"url":"https://arxiv.org/abs/2506.18701","region":"t38","hemi":"wm","lobe":"temporal","color":"#9a4f66","indegree":3,"alsoIn":[],"x":0.3411,"y":0.5911,"text":"Matrix-Game addresses controllable game-world generation by training a 17B-parameter image-to-world diffusion model in two stages: large-scale unlabeled pretraining for environment understanding, followed by action-labeled training for interactive generation. The authors curate Matrix-Game-MC (2,700+ hours unlabeled and 1,000+ hours action-labeled Minecraft clips with fine-grained keyboard/mouse annotations) and introduce GameWorld Score, an 8-dimension benchmark covering visual quality, temporal quality, action controllability, and physical rule understanding. The model is conditioned on a reference image, motion context, and user actions, generating videos autoregressively for long horizons. It outperforms open-source baselines Oasis and MineWorld across all metrics, especially controllability and physical consistency, and is strongly preferred in double-blind human evaluation."},{"id":"arxiv:2503.15168","kind":"paper","n":527,"label":"World Models in Artificial Intelligence: Sensing, Learning, and Reasoning Like a Child","authors":"","year":null,"url":"https://arxiv.org/abs/2503.15168","region":"t7","hemi":"llm","lobe":"occipital","color":"#adb067","indegree":15,"alsoIn":[],"x":0.6734,"y":0.6168,"text":"This position paper argues that modern AI, especially LLMs, remains confined to statistical pattern recognition and lacks genuine reasoning, causal understanding, and common sense \u2014 capabilities even young children develop. Drawing on Piaget's constructivist theory of cognitive development (its four stages, schemas, assimilation/accommodation), the authors propose building structured, adaptive, interpretable World Models that learn progressively through perception, representation, reasoning, and generalization stages. They identify six interdependent research areas \u2014 physics-informed ML and embodied AI, neurosymbolic systems, causal inference, open-world/continual learning, human-in-the-loop AI, and trustworthy/responsible AI \u2014 as essential to enabling true reasoning. The central thesis is that bridging Piagetian constructivism with modern AI is a feasible path beyond pattern recognition toward reasoning, adaptation, and generalization."},{"id":"arxiv:2510.19788","kind":"paper","n":528,"label":"Benchmarking World-Model Learning (WorldTest / AutumnBench)","authors":"Archana Warrier, Dat Nguyen, Michelangelo Naim, et al.","year":2025,"url":"https://arxiv.org/abs/2510.19788","region":"t17","hemi":"llm","lobe":"parietal","color":"#4f9a79","indegree":1,"alsoIn":[],"x":0.6809,"y":0.5941,"text":"Current world-model evaluations only test properties measurable from observed trajectories (e.g., next-frame prediction, task return) and don't assess whether a learned model supports diverse queries about an environment. The paper proposes WorldTest, a behavior-based, reward-free, representation-agnostic two-phase protocol (reward-free interaction then test on derived challenge environments) that probes environment-level queries, and instantiates it as AutumnBench: 43 interactive grid-world environments and 129 tasks across three families (masked frame prediction, change detection, planning). Experiments with 517 human participants and five frontier reasoning models show humans substantially outperform all models across every environment and task type. The gap is attributed to differences in exploration (use of resets/no-ops for hypothesis testing) and belief updating rather than to compute scaling."},{"id":"nature.com:ef956cce5b54","kind":"paper","n":529,"label":"Geometry of abstract learned knowledge in the hippocampus","authors":"Edward H. Nieh, Manuel Schottdorf, Nicolas W. Freeman, Ryan J. Low, Sam Lewallen","year":2021,"url":"https://www.nature.com/articles/s41586-021-03652-7","region":"t28","hemi":"llm","lobe":"frontal","color":"#7667b0","indegree":7,"alsoIn":[],"x":0.5523,"y":0.4964,"text":"Using two-photon calcium imaging of CA1 neurons in the dorsal hippocampus of mice performing an accumulating-towers evidence-accumulation decision task in virtual reality, the authors show that individual neurons jointly encode an abstract learned variable (accumulated evidence) together with the physical variable of spatial position. Nonlinear dimensionality reduction (the MIND algorithm) revealed that population activity lies on a low-dimensional manifold described by roughly four to six latent variables, on which physical and abstract variables are mapped together in an orderly geometric arrangement. This conjoined cognitive map is similar across mice, and single-trial sequential CA1 activity predicts the animal's upcoming choice and is explained by the manifold. The finding suggests the hippocampus performs a general computation: building task-specific low-dimensional manifolds that geometrically represent learned knowledge."},{"id":"arxiv:2306.10754","kind":"paper","n":530,"label":"Collaborative Optimization of Multi-microgrids System with Shared Energy Storage Based on Multi-agent Stochastic Game and Reinforcement Learning","authors":"","year":null,"url":"https://arxiv.org/abs/2306.10754","region":"t4","hemi":"wm","lobe":"frontal","color":"#9a7c4f","indegree":3,"alsoIn":[],"x":0.2807,"y":0.1664,"text":"This paper addresses the collaborative optimization of multi-microgrid (MMG) systems with shared energy storage (SES) under partially observable dynamic stochastic game conditions, considering nonlinear equipment characteristics and privacy protection. A Mixed-Attention mechanism is proposed to fit nonlinear conditions of energy conversion equipment, and MA-SAC and MA-WoLF-PHC algorithms are developed for energy management and bidding respectively. Tested on operation data from an MMG system in Northwest China, the framework reduces main grid energy fluctuations by 1746.5kW in 24 hours and achieves 16.21% cost reduction."},{"id":"arxiv:2102.11107","kind":"paper","n":531,"label":"Toward Causal Representation Learning","authors":"Bernhard Sch\u00f6lkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalc","year":2021,"url":"https://arxiv.org/abs/2102.11107","region":"t6","hemi":"llm","lobe":"temporal","color":"#b0a967","indegree":3,"alsoIn":[],"x":0.648,"y":0.7155,"text":"This position/review paper bridges machine learning and graphical causality, arguing that current ML's reliance on i.i.d. pattern recognition explains its failures at out-of-distribution generalization, transfer, and robustness, all of which causal modeling addresses by representing the mechanisms behind statistical dependences. It reviews structural causal models (SCMs), the Independent Causal Mechanisms (ICM) principle, and causal discovery, then formalizes the open problem of causal representation learning: discovering high-level causal variables from low-level (e.g. pixel) observations. It introduces the Sparse Mechanism Shift (SMS) hypothesis as a consequence of ICM and recasts ML practices (semi-supervised learning, data augmentation, self-supervision, adversarial robustness, RL/world models) in causal terms, proposing modular, recomposable architectures as a path to more versatile AI."},{"id":"arxiv:2411.13157","kind":"paper","n":532,"label":"Closer Look at Efficient Inference Methods: A Survey of Speculative Decoding","authors":"Hyun Ryu, Eric Kim","year":2024,"url":"https://arxiv.org/abs/2411.13157","region":"t9","hemi":"llm","lobe":"parietal","color":"#97b067","indegree":6,"alsoIn":[],"x":0.8291,"y":0.7731,"text":"This survey reviews speculative decoding, a technique to accelerate autoregressive LLM inference by splitting generation into a fast parallel drafting phase (small model) and a verification phase (large target model), which alleviates the memory-access bottleneck of sequential decoding. The authors propose a taxonomy dividing methods into model-centric implementations (improving draft quality/efficiency via independent vs. dependent drafter architectures like Medusa, Hydra, EAGLE, early-exit/LayerSkip) and draft-centric implementations (refining the candidate pool via probability-based, search-optimization, tree/graph-based, and hybrid/adaptive approaches like EAGLE-2, OPT-Tree, GSD). It then surveys real-world deployment challenges: throughput, long-context generation, model parallelism, hardware limitation, and generalizability. The paper concludes that speculative decoding is promising but that addressing these deployment limitations is critical for practical use."},{"id":"arxiv:2001.11841","kind":"paper","n":533,"label":"Learning Perception and Planning with Deep Active Inference","authors":"","year":null,"url":"https://arxiv.org/abs/2001.11841","region":"t14","hemi":"llm","lobe":"temporal","color":"#67b06e","indegree":6,"alsoIn":[],"x":0.5942,"y":0.892,"text":"Active inference, a neuroscience-derived framework where agents minimize expected free energy, has been limited to predefined discrete state spaces. This paper proposes using deep neural networks to learn the state space, transition model, and approximate posterior distributions, enabling active inference with learned representations. The method is demonstrated on the Mountain Car control task, where the agent plans by Monte Carlo sampling of trajectories and selects policies minimizing expected free energy, balancing goal-directed behavior and uncertainty resolution via a tunable parameter \u03c1."},{"id":"arxiv:2412.17254","kind":"paper","n":534,"label":"ART\u2022V: aside \u2014 Enhancing Long Video Generation Consistency without Tuning","authors":"Xingyao Li et al.","year":2024,"url":"https://arxiv.org/abs/2412.17254","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":1,"alsoIn":[],"x":0.3844,"y":0.3913,"text":"The paper tackles frame-to-frame inconsistency (abrupt object appearance/disappearance, background and scene changes) in training-free long video generation using diffusion models. It observes that temporal attention scores over-concentrate on the diagonal for inconsistent frames and that inconsistency corresponds to abnormally high-frequency components; based on this it proposes TiARA, which reweights the temporal attention matrix adaptively using motion intensity derived from the Discrete Short-Time Fourier Transform (DSTFT), plus PromptBlend, a prompt-alignment-and-interpolation pipeline for multi-prompt transitions. TiARA is a plug-in for FIFO-Diffusion, FreeNoise, StreamingT2V and CogVideoX, and a theoretical guarantee (Theorem 4.1) bounds the reduction of inconsistency error. Extensive experiments on VBench/EvalCrafter metrics show consistent gains in subject/background consistency and temporal quality with no loss of image quality."},{"id":"arxiv:2606.10749","kind":"paper","n":535,"label":"Toward Secure LLM Agents: Threat Surfaces, Attacks, Defenses, and Evaluation","authors":"(arXiv 2606.10749)","year":2026,"url":"https://arxiv.org/abs/2606.10749","region":"t8","hemi":"llm","lobe":"frontal","color":"#a2b067","indegree":2,"alsoIn":[],"x":0.6104,"y":0.9447,"text":"This paper is a systematic survey of 247 papers (2023\u2013April 2026) on LLM agent security, reframing the problem from prompt-level model safety to a software/systems security problem around the agentic loop. It introduces a lifecycle-based, systems-oriented framework modeling an agent as a tuple A=\u27e8I,P,D,T,M,O,C\u27e9 (input, planning, decision, tool execution, memory, output, coordination) shaped by information flow, delegated authority, and persistent state, and answers four RQs on scope/modeling, threat surfaces, defenses, and evaluation. It finds prompt injection and tool-mediated control-flow hijacking still dominate, while persistent state corruption and multi-agent propagation are rising concerns; defenses provide useful building blocks but lack a stable compositional stack; and benchmarks underrepresent long-horizon, stateful, and deployment-sensitive risks. It argues secure agents need explicit trust boundaries, privilege control, provenance-aware state management, and deployment-realistic evaluation."},{"id":"arxiv:2407.02485","kind":"paper","n":536,"label":"RankRAG: Unifying Context Ranking with Retrieval-Augmented Generation in LLMs","authors":"Yue Yu et al.","year":2024,"url":"https://arxiv.org/abs/2407.02485","region":"t31","hemi":"llm","lobe":"occipital","color":"#9767b0","indegree":2,"alsoIn":[],"x":0.8559,"y":0.6316,"text":"Current RAG pipelines suffer from a trade-off in selecting top-k contexts: small k hurts recall while large k introduces noise, and separate expert ranking models generalize poorly. The paper proposes RankRAG, a two-stage instruction-tuning framework that fine-tunes a single LLM to both rank retrieved contexts and generate answers, using a retrieve-rerank-generate inference pipeline. Adding only ~50k ranking examples (~1% of MS MARCO) lets RankRAG outperform dedicated rankers trained on 10x more data, and Llama3-RankRAG-8B/70B beat Llama3-ChatQA-1.5 and GPT-4 across nine knowledge-intensive benchmarks. It also matches ~98% of GPT-4 performance on five biomedical RAG tasks (Mirage) with no in-domain tuning, showing strong domain generalization."},{"id":"arxiv:2503.02854","kind":"paper","n":537,"label":"(How) Do Language Models Track State?","authors":"Li, Guo & Andreas","year":2025,"url":"https://arxiv.org/abs/2503.02854","region":"t16","hemi":"llm","lobe":"frontal","color":"#67b084","indegree":2,"alsoIn":[],"x":0.7527,"y":0.6624,"text":"The paper investigates the mechanisms transformer LMs use to track latent state, using permutation composition (computing object order after a sequence of swaps, in S3 and S5) as an NC1-complete model system for general state tracking. Training/fine-tuning Pythia-160M and GPT-2 models on length-100 permutation sequences and analyzing them with activation patching, linear probing, attention analysis, and training dynamics, they find LMs consistently learn one of two mechanisms: an 'associative algorithm' (AA) that hierarchically composes action subsequences in parallel across layers, or a 'parity-associative algorithm' (PAA) that first computes a parity heuristic in early layers then an associative residual. AA models generalize better and converge faster; the chosen mechanism is a deterministic function of architecture, size, initialization, and data order, and can be steered via intermediate tasks (topic-modeling pretraining \u2192 AA; parity-prediction curriculum \u2192 PAA). No evidence is found for step-by-step sequential simulation or fully parallel composition despite both being theoretically implementable."},{"id":"arxiv:2406.03520","kind":"paper","n":538,"label":"VideoPhy: Evaluating Physical Commonsense for Video Generation","authors":"Bansal, Lin, Peng et al. (UCLA & Google)","year":2024,"url":"https://arxiv.org/abs/2406.03520","region":"t29","hemi":"wm","lobe":"parietal","color":"#6a4f9a","indegree":5,"alsoIn":[],"x":0.4492,"y":0.8136,"text":"The paper introduces VideoPhy, a benchmark of 688 human-verified captions covering solid-solid, solid-fluid, and fluid-fluid material interactions, used to test whether text-to-video models generate videos that obey physical commonsense and adhere to the prompt. Across 12 open and closed T2V models (11,330 generated videos, 36,500 human annotations), the authors find all models perform poorly, with the best (CogVideoX-5B) producing videos that satisfy both semantic adherence and physical commonsense for only 39.6% of prompts and every other model scoring below 20%. They also release VideoCon-Physics, a finetuned 7B video-language auto-evaluator that judges semantic adherence and physical commonsense more reliably than GPT-4-Vision and Gemini-1.5-Pro. The work concludes that current T2V models are far from being general-purpose physical world simulators, especially for solid-solid interactions."},{"id":"arxiv:2604.22748","kind":"paper","n":539,"label":"Agentic World Modeling: Foundations, Capabilities, Laws, and Beyond","authors":"Matrix-Agent team (see arXiv 2604.22748)","year":2026,"url":"https://arxiv.org/abs/2604.22748","region":"t38","hemi":"llm","lobe":"temporal","color":"#9a4f66","indegree":1,"alsoIn":[],"x":0.6521,"y":0.6902,"text":"This position-driven survey addresses conceptual fragmentation in how different research communities use the term 'world model' by proposing a two-axis 'levels\u00d7laws' taxonomy. The first axis defines three capability levels \u2014 L1 Predictor (one-step local transition operators), L2 Simulator (multi-step, action-conditioned, constraint-respecting rollouts), and L3 Evolver (autonomous evidence-driven revision of the model itself) \u2014 each with testable boundary conditions; the second axis defines four governing-law regimes (physical, digital, social, scientific) that determine the constraints a world model must satisfy. Synthesizing over 400 works and summarizing more than 100 representative systems across model-based RL, video generation, web/GUI agents, multi-agent social simulation, and AI-driven scientific discovery, it analyzes methods, failure modes, and evaluation practices per level\u2013regime pair. It contributes decision-centric evaluation principles, a Minimal Reproducible Evaluation Package (MREP), architectural guidance, and an open-problems roadmap toward symbolic, self-revising world models."},{"id":"arxiv:2101.03961","kind":"paper","n":540,"label":"Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity","authors":"William Fedus, Barret Zoph, Noam Shazeer","year":2021,"url":"https://arxiv.org/abs/2101.03961","region":"t16","hemi":"llm","lobe":"frontal","color":"#67b084","indegree":2,"alsoIn":[],"x":0.7437,"y":0.4324,"text":"Dense Transformers reuse all parameters for every input, making scale extremely compute-intensive. The paper introduces the Switch Transformer, a simplified Mixture-of-Experts architecture that routes each token to exactly one expert (k=1), decoupling parameter count from FLOPs-per-token, plus training techniques (selective float32 precision in the router, reduced initialization scale, expert dropout) that stabilize training even in bfloat16. FLOP-matched against T5, Switch achieves up to 7x faster pre-training, distills back into dense models preserving ~30% of quality gains, improves all 101 languages in multilingual training, and scales to a 1.6T-parameter model that is 4x faster to a fixed perplexity than T5-XXL."},{"id":"arxiv:2310.09615","kind":"paper","n":541,"label":"STORM: Efficient Stochastic Transformer based World Models for Reinforcement Learning","authors":"Zhang et al.","year":2023,"url":"https://arxiv.org/abs/2310.09615","region":"t26","hemi":"wm","lobe":"temporal","color":"#4f579a","indegree":4,"alsoIn":[],"x":0.1403,"y":0.4398,"text":"Model-based RL agents trained in a learned 'imagination' world model suffer from accumulated autoregressive prediction errors that cause agents to chase virtual goals; existing Transformer world models (IRIS, TWM) also train slowly and fail to beat the GRU-based DreamerV3. STORM combines a GPT-like Transformer sequence model with a categorical-VAE stochastic image encoder, fusing each latent observation and action into a single token to inject beneficial random noise while keeping training parallelizable. On the Atari 100k benchmark STORM reaches a mean human-normalized score of 126.7% (median 58.4%), a new record among methods without lookahead search, while training in only 4.3 hours on a single RTX 3090 for 1.85 hours of simulated interaction."},{"id":"arxiv:2402.10644","kind":"paper","n":542,"label":"Linear Transformers with Learnable Kernel Functions are Better In-Context Models","authors":"Yaroslav Aksenov et al.","year":2024,"url":"https://arxiv.org/abs/2402.10644","region":"t16","hemi":"llm","lobe":"frontal","color":"#67b084","indegree":4,"alsoIn":[],"x":0.7841,"y":0.3567,"text":"Linear Transformers and State Space Models scale subquadratically but underperform Transformers on In-Context Learning, especially the Multi-Query Associative Recall (MQAR) task at long sequences and small model sizes. The paper analyzes the Based model's Taylor-expansion kernel and identifies that its similarity function has a fixed minimum of 0.5 and an awkward optimum at q^T k = -1, preventing near-zero attention scores. The authors propose ReBased, which replaces the kernel with a learnable quadratic phi(x)=x^2 wrapped in an affine transformation (learnable scale gamma and shift beta) plus Layer Normalization before the kernel. ReBased outperforms Based on MQAR across lengths/sizes and yields lower perplexity on the Pile and better few-shot results, narrowing the gap to full attention."},{"id":"arxiv:2306.02572","kind":"paper","n":543,"label":"Introduction to Latent Variable Energy-Based Models: A Path Towards Autonomous Machine Intelligence","authors":"Anna Dawid, Yann LeCun","year":2023,"url":"https://arxiv.org/abs/2306.02572","region":"t14","hemi":"llm","lobe":"temporal","color":"#67b06e","indegree":1,"alsoIn":[],"x":0.6705,"y":0.5734,"text":"These lecture notes provide a pedagogical introduction to energy-based models (EBMs) and latent-variable EBMs as the conceptual foundation for Yann LeCun's proposed architecture for autonomous machine intelligence. The authors argue that current ML (supervised and reinforcement learning) is sample-inefficient, brittle, and lacks world models and 'common sense,' then motivate replacing intractable high-dimensional probabilistic models with EBMs trained via contrastive vs. regularized methods. They build up to the joint embedding predictive architecture (JEPA) and its hierarchical extension (H-JEPA), which predict in representation space using latent variables to handle uncertainty/multimodality and stacked levels to enable multi-timescale hierarchical planning. The paper is a tutorial/position synthesis of LeCun's 2022 'A Path Towards Autonomous Machine Intelligence' rather than an empirical study."},{"id":"arxiv:2406.01506","kind":"paper","n":544,"label":"The Geometry of Categorical and Hierarchical Concepts in Large Language Models","authors":"Park, Choe, Jiang & Veitch","year":2024,"url":"https://arxiv.org/abs/2406.01506","region":"t6","hemi":"llm","lobe":"temporal","color":"#b0a967","indegree":2,"alsoIn":[],"x":0.651,"y":0.4455,"text":"The paper extends the linear representation hypothesis from binary concepts with natural contrasts (e.g., male\u21d2female) to general binary features without contrasts (e.g., is-animal), showing how to assign magnitudes so features become vectors rather than mere directions. From this it formalizes categorical concepts as polytopes (simplices for 'natural' concepts) and proves that semantic hierarchy is encoded geometrically as orthogonality between concept representations. The authors validate the theory on Gemma-2B and LLaMA-3-8B by estimating vector representations for 900+ WordNet noun/verb synsets via LDA, finding that cosine similarities and child-parent orthogonality match the predicted structure while shuffled-embedding controls do not."},{"id":"arxiv:2312.10997","kind":"paper","n":545,"label":"Retrieval-Augmented Generation for Large Language Models: A Survey","authors":"Yunfan Gao et al.","year":2023,"url":"https://arxiv.org/abs/2312.10997","region":"t31","hemi":"llm","lobe":"occipital","color":"#9767b0","indegree":5,"alsoIn":[],"x":0.8843,"y":0.5914,"text":"This survey systematically reviews Retrieval-Augmented Generation (RAG), which mitigates LLM hallucination, outdated knowledge, and opaque reasoning by retrieving relevant document chunks from external knowledge bases and injecting them into the prompt. It organizes over 100 RAG studies into three evolving paradigms\u2014Naive RAG (retrieve-read), Advanced RAG (pre-/post-retrieval optimization), and Modular RAG (reconfigurable modules with iterative, recursive, and adaptive retrieval)\u2014and dissects the three core pillars of retrieval, generation, and augmentation. It further compiles RAG's downstream tasks, datasets, evaluation targets/metrics, benchmarks (RGB, RECALL, CRUD) and tools (RAGAS, ARES, TruLens), and outlines open challenges including RAG-vs-long-context, robustness, hybrid RAG+fine-tuning, scaling laws, and multimodal extension."},{"id":"arxiv:1911.03537","kind":"paper","n":546,"label":"A filamentary cascade model of the inertial range","authors":"","year":null,"url":"https://arxiv.org/abs/1911.03537","region":"t36","hemi":"llm","lobe":"frontal","color":"#b06793","indegree":4,"alsoIn":[],"x":0.6495,"y":0.5161,"text":"The paper proposes a toy model of the inertial range of high-Reynolds-number turbulence built from a binary cascade of slender helical vortex filaments, where each filament splits into two daughter filaments carrying differing circulation fractions, stretching factors, and scaling ratios \u2014 in effect two simultaneous Richardson\u2013Kolmogorov cascades, neither having the K41 1/3 velocity exponent. Vorticity volume is conserved (space-filling, not fractal), and the only dynamical constraint imposed is Kolmogorov's four-fifths law (\u03b63 = 1). By varying essentially one parameter (\u03b20, with s1 fixed by the four-fifths law and s0 tuned), the structure functions \u03b6p match the She\u2013Leveque formula extremely well for 1 \u2264 p \u2264 10, with the two distinct scaling factors producing the observed nonlinear (intermittent) dependence of \u03b6p on p. Treating a single cascade as an initial-value problem and letting it proceed indefinitely lets the authors compute the timing of energy delivery to vanishingly small scales and hence inviscid energy decay."},{"id":"arxiv:2301.08861","kind":"paper","n":547,"label":"Data-Driven Distributionally Robust Scheduling of Community Integrated Energy Systems with Uncertain Renewable Generations Considering Integrated Demand Response","authors":"","year":null,"url":"https://arxiv.org/abs/2301.08861","region":"t4","hemi":"wm","lobe":"frontal","color":"#9a7c4f","indegree":4,"alsoIn":[],"x":0.2159,"y":0.1363,"text":"This paper addresses the optimal scheduling of Community Integrated Energy Systems (CIES) under renewable generation uncertainty by proposing a data-driven two-stage distributionally robust optimization (DRO) model. The method uses a WGAN-GP for scenario generation, a comprehensive norm (1-norm and \u221e-norm) as the probability distribution ambiguity set, and an integrated demand response (IDR) mechanism that accounts for human thermal comfort ambiguity and building thermal inertia. Simulated on an actual CIES in North China, the proposed DRO balances economy and robustness better than stochastic programming and robust optimization, and outperforms moment-based and Wasserstein-based DRO methods in operating cost, curtailment rate, and computational efficiency."},{"id":"arxiv:2111.06377","kind":"paper","n":548,"label":"Masked Autoencoders Are Scalable Vision Learners","authors":"Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll\u00e1r, Ross Girshick","year":2022,"url":"https://arxiv.org/abs/2111.06377","region":"t23","hemi":"wm","lobe":"occipital","color":"#4f799a","indegree":0,"alsoIn":[],"x":0.3591,"y":0.2484,"text":"The paper introduces Masked Autoencoders (MAE), a simple self-supervised pre-training method for vision that masks a high proportion of random image patches (e.g., 75%) and reconstructs the missing pixels. It uses an asymmetric encoder-decoder where the encoder processes only the visible (~25%) patches and a lightweight decoder reconstructs from latent representations plus mask tokens, accelerating training 3x or more while reducing memory. A vanilla ViT-Huge fine-tuned on ImageNet-1K reaches 87.8% top-1 accuracy, the best among ImageNet-1K-only methods, and the pre-training transfers better than supervised pre-training on detection, segmentation, and classification with strong scaling behavior."},{"id":"arxiv:2311.17593","kind":"paper","n":549,"label":"LanGWM: Language Grounded World Model","authors":"Poudel, Pandya, Zhang & Cipolla","year":2023,"url":"https://arxiv.org/abs/2311.17593","region":"t17","hemi":"llm","lobe":"parietal","color":"#4f9a79","indegree":0,"alsoIn":[],"x":0.7258,"y":0.6178,"text":"State-of-the-art image-based RL models struggle with out-of-distribution (OoD) generalization on visual control tasks. LanGWM improves world-model (model-based RL) state abstraction by explicitly grounding visual features in language: it masks the bounding boxes of a few objects in the observation, supplies templated text descriptions of those masked objects, and reconstructs the masked regions (depth) via a transformer-based masked autoencoder, then learns the navigation policy in imagined rollouts. On the iGibson 1.0 PointGoal navigation OoD benchmark at 100K interaction steps, LanGWM achieves the best results among comparable RL methods (6.8% average success rate vs 5.5% for DreamerV2+DA and 4.7% for CURL). Ablations show both object masking and language descriptions are necessary and complementary."},{"id":"arxiv:2601.01885","kind":"paper","n":550,"label":"Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model Agents","authors":"Yi Yu, Liuyi Yao, Yuexiang Xie, Qingquan Tan, Jiaqi Feng, Yaliang Li, Libing Wu","year":2026,"url":"https://arxiv.org/abs/2601.01885","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":6,"alsoIn":[],"x":0.6145,"y":0.9537,"text":"LLM agents are limited by finite context windows, and existing methods manage long-term memory (LTM) and short-term memory (STM) as separate components governed by heuristics or auxiliary controllers. This paper proposes AgeMem, a unified framework that exposes six memory operations (Add/Update/Delete for LTM; Retrieve/Summary/Filter for STM) as tool-based actions integrated directly into the agent's policy, trained via a three-stage progressive RL curriculum and a step-wise GRPO that broadcasts the terminal task reward back to all memory decisions. Across five long-horizon benchmarks, AgeMem raises average task performance (e.g., 41.96% on Qwen2.5-7B and 54.31% on Qwen3-4B), improves stored long-term memory quality (MQ up to 0.605), and reduces prompt token usage versus RAG-based variants. Ablations show RL training and the multi-component reward each contribute substantial gains over no-RL and answer-only baselines."},{"id":"arxiv:2005.11401","kind":"paper","n":551,"label":"Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks","authors":"Patrick Lewis et al.","year":2020,"url":"https://arxiv.org/abs/2005.11401","region":"t31","hemi":"llm","lobe":"occipital","color":"#9767b0","indegree":4,"alsoIn":[],"x":0.8709,"y":0.6418,"text":"Large pre-trained parametric LMs store knowledge but can't easily revise it, cite provenance, or reliably manipulate it on knowledge-intensive tasks. The paper introduces Retrieval-Augmented Generation (RAG), a general fine-tuning recipe pairing a pre-trained seq2seq generator (BART-large) as parametric memory with a dense vector index of 21M Wikipedia passages accessed via a DPR retriever as non-parametric memory, trained end-to-end by marginalizing over retrieved latent documents. Two variants are compared: RAG-Sequence (one document for the whole output) and RAG-Token (per-token document choice). RAG sets state of the art on three open-domain QA tasks and generates more factual, specific, and diverse text than a BART baseline on generation tasks."},{"id":"arxiv:2305.17888","kind":"paper","n":552,"label":"LLM-QAT: Data-Free Quantization Aware Training for Large Language Models","authors":"Zechun Liu, Barlas Oguz, Changsheng Zhao, et al.","year":2023,"url":"https://arxiv.org/abs/2305.17888","region":"t37","hemi":"llm","lobe":"parietal","color":"#b06788","indegree":7,"alsoIn":[],"x":0.8813,"y":0.3557,"text":"Post-training quantization (PTQ) methods for LLMs work well at 8-bit but break down below 8 bits. This paper introduces LLM-QAT, the first quantization-aware training method for LLMs, using a data-free knowledge distillation approach that fine-tunes the quantized student on data generated by the pre-trained model itself (with soft-label logit distillation), and additionally quantizes the KV cache alongside weights and activations. On LLaMA-7B/13B/30B down to 4-bit weights/KV cache and 6-8-bit activations, LLM-QAT substantially outperforms PTQ baselines (RTN, GPTQ, SmoothQuant), especially in low-bit settings, and larger quantized models beat smaller FP16 models of similar size."},{"id":"arxiv:2510.08531","kind":"paper","n":553,"label":"SpatialLadder: Progressive Training for Spatial Reasoning in Vision-Language Models","authors":"Hongxing Li et al.","year":2025,"url":"https://arxiv.org/abs/2510.08531","region":"t12","hemi":"llm","lobe":"frontal","color":"#5e9a4f","indegree":6,"alsoIn":[],"x":0.6375,"y":0.1493,"text":"VLMs struggle with spatial reasoning because existing methods attempt to learn it directly without first establishing perceptual foundations. The authors introduce SpatialLadder-26k, a 26,610-sample multimodal dataset spanning object localization, single-image, multi-view, and video spatial reasoning, and a three-stage progressive training framework: (1) perceptual grounding via object localization, (2) spatial understanding via multi-dimensional SFT tasks, (3) complex reasoning via GRPO with verifiable rewards. The resulting 3B model achieves 62.3% overall accuracy on in-domain benchmarks, surpassing GPT-4o (41.5%) and Gemini-2.0-Flash (52.2%), while maintaining 7.2% improvement on out-of-domain benchmarks."},{"id":"arxiv:1803.07616","kind":"paper","n":554,"label":"IntPhys: A Framework and Benchmark for Visual Intuitive Physics Reasoning","authors":"Riochet et al.","year":2018,"url":"https://arxiv.org/abs/1803.07616","region":"t6","hemi":"llm","lobe":"temporal","color":"#b0a967","indegree":3,"alsoIn":[],"x":0.6435,"y":0.3485,"text":"IntPhys 2019 introduces a benchmark, inspired by the violation-of-expectation (VOE) paradigm from infant cognition, that diagnoses a vision system's intuitive physics understanding by requiring it to output a scalar plausibility score per video and discriminate well-matched possible vs. impossible events. Stimuli are procedurally generated in Unreal Engine 4 across three concept blocks (object permanence O1, shape constancy O2, spatio-temporal continuity O3), each varied along visibility, motion complexity, and object count, with pixel-matched quadruplets to remove low-level biases. Two unsupervised baselines (a CNN encoder-decoder and a conditional GAN) are trained only on possible videos via future semantic-mask prediction and scored by aggregating per-frame prediction error (minimum over frames). Both models perform above chance when violations are visible but collapse to near-chance when occluded, far below humans, showing next-frame prediction lacks the memory to handle occlusion."},{"id":"arxiv:2305.16653","kind":"paper","n":555,"label":"AdaPlanner: Adaptive Planning from Feedback with Language Models","authors":"Sun et al.","year":2023,"url":"https://arxiv.org/abs/2305.16653","region":"t19","hemi":"llm","lobe":"occipital","color":"#67b0a5","indegree":1,"alsoIn":[],"x":0.5623,"y":0.9412,"text":"Existing LLM agents for sequential decision-making either act greedily without planning or follow static plans that cannot adapt to environment feedback, degrading on complex, long-horizon tasks. AdaPlanner is a closed-loop method where one LLM acts as both planner and refiner, using in-plan refinement (an ask_LLM() action that parses aligned observations) and out-of-plan refinement (revising the whole plan via a refine-then-resume mechanism), plus code-style prompts to curb hallucination and a skill-discovery memory to reuse successful plans as few-shot exemplars. On ALFWorld it reaches 91.79% success and on MiniWoB++ (feedback tasks) 91.11%, beating SOTA baselines by 3.73% and 4.11% while using 2x and 600x fewer samples respectively. Ablations show the code interface and skill discovery are each critical, with large success-rate drops when removed."},{"id":"arxiv:2507.02768","kind":"paper","n":556,"label":"DeSTA2.5-Audio: Toward General-Purpose Large Audio Language Model with Self-Generated Cross-Modal Alignment","authors":"Ke-Han Lu et al.","year":2025,"url":"https://arxiv.org/abs/2507.02768","region":"t11","hemi":"llm","lobe":"occipital","color":"#81b067","indegree":3,"alsoIn":[],"x":0.835,"y":0.2647,"text":"Existing Large Audio Language Models (LALMs) suffer catastrophic forgetting of the backbone LLM's original text abilities because they train on targets generated by foreign LLMs or humans, introducing a distribution mismatch. DeSTA2.5-Audio proposes a self-generated cross-modal alignment strategy (DeSTA) where the backbone LLM itself produces the training targets from structured audio descriptions plus sampled prompts, eliminating stylistic shift and preserving instruction-following. The authors build DeSTA-AQA5M, a 5M-sample task-agnostic dataset from 7,000 hours of audio across 50 datasets (speech, sound, music), and train a frozen-LLM/frozen-encoder model with only a Q-Former modality adapter (131M trainable of 8.8B params). Despite using only 7,000 hours vs. Qwen2-Audio-Instruct's 510,000 hours, it reaches SOTA or competitive results on Dynamic-SUPERB, MMAU, SAKURA, Speech-IFEval, and VoiceBench."},{"id":"arxiv:2605.12957","kind":"paper","n":557,"label":"GTA: Advancing Image-to-3D World Generation via Geometry Then Appearance Video Diffusion","authors":"","year":null,"url":"https://arxiv.org/abs/2605.12957","region":"t15","hemi":"wm","lobe":"occipital","color":"#4f9a62","indegree":7,"alsoIn":[],"x":0.3951,"y":0.6002,"text":"Existing image-to-3D world generation methods are appearance-centric or jointly predict geometry and appearance, causing unreliable scene structure and poor cross-view consistency. This paper proposes GTA, a two-stage Geometry-Then-Appearance framework using two dedicated video diffusion models (built on CogVideoX-5B): one first synthesizes a multi-view depth video via depth-based warping, then a second synthesizes RGB appearance conditioned on that geometry. A random latent shuffle training strategy and a test-time scaling scheme further improve cross-view consistency and perceptual quality. GTA outperforms SOTA on DL3DV (PSNR 17.47 vs 16.15 for Gen3C) and held-out RealEstate10K, works as a plug-in post-hoc enhancer, and is highly data-efficient."},{"id":"arxiv:2305.11169","kind":"paper","n":558,"label":"Emergent Representations of Program Semantics in Language Models Trained on Programs","authors":"Jin & Rinard","year":2024,"url":"https://arxiv.org/abs/2305.11169","region":"t1","hemi":"llm","lobe":"parietal","color":"#b07267","indegree":5,"alsoIn":[],"x":0.7557,"y":0.7531,"text":"The paper investigates whether a language model trained purely on next-token prediction over code learns the formal semantics of programs, not just surface statistics. The authors train a 350M-parameter CodeGen Transformer on 500,000 synthetic Karel grid-world programs (each preceded by 5 input-output example grids) and use small probing classifiers to extract unobserved intermediate program states (the abstract interpretation of position, direction, obstacle) from the LM's hidden states. They find that probe-extractable semantic content emerges during a distinct 'semantics acquisition' training phase and is strongly linearly correlated with the LM's program-synthesis accuracy, and they introduce a 'semantic probing intervention' baseline that rules out the alternative that the probe (rather than the LM) is doing the semantic work. The results reject the hypothesis that next-token-trained LMs of code cannot model formal program semantics."},{"id":"arxiv:2409.10644","kind":"paper","n":559,"label":"Improving Multi-candidate Speculative Decoding","authors":"","year":null,"url":"https://arxiv.org/abs/2409.10644","region":"t9","hemi":"llm","lobe":"parietal","color":"#97b067","indegree":1,"alsoIn":[],"x":0.8499,"y":0.8313,"text":"The paper tackles inefficiencies in Multi-Candidate Speculative Decoding (MCSD), where a small draft model proposes multiple candidate tokens verified in parallel by a larger target model but suffers from draft/target distribution mismatch, static draft length, and costly topology-aware mask construction. It introduces three techniques: target-model-initialized multi-candidate generation (sampling multiple tokens from the target model to seed the token tree), a dynamic sliced topology-aware causal mask for adjustable draft length without mask rebuilding, and MLP-based early-stop decision models. The static target-initialized variant with a (2,4,3,1,1) configuration achieves up to 27.5% speedup over the MCSD baseline (max 1.90x over vanilla on MT-Bench) using Llama 2-7B as target and Llama-68M as draft, raising acceptance rate substantially. However, the full dynamic framework with decision models fails to beat the static method, and target initialization with more than one token degrades output quality (e.g., MT-Bench drops from 6.29 to 5.07)."},{"id":"arxiv:2311.17042","kind":"paper","n":560,"label":"Adversarial Diffusion Distillation","authors":"Axel Sauer, Dominik Lorenz, Andreas Blattmann, Robin Rombach","year":2023,"url":"https://arxiv.org/abs/2311.17042","region":"t32","hemi":"wm","lobe":"frontal","color":"#8b4f9a","indegree":7,"alsoIn":[],"x":0.4049,"y":0.1122,"text":"Adversarial Diffusion Distillation (ADD) is a training method that distills large pretrained image diffusion models (Stable Diffusion, SDXL) into generators that produce high-fidelity 512x512 images in just 1\u20134 sampling steps. It combines an adversarial loss (using a frozen ViT feature network with lightweight discriminator heads) with a score-distillation loss from a frozen diffusion teacher, initializing the student from pretrained diffusion weights and avoiding classifier-free guidance at inference. In a single step ADD outperforms GANs and Latent Consistency Models, and with four steps ADD-XL beats its teacher SDXL-Base (50 steps) in human preference studies. ADD-M (SD1.5 backbone) also achieves the best FID/CLIP among distillation methods, enabling real-time single-step generation with foundation models."},{"id":"arxiv:2508.16745","kind":"paper","n":561,"label":"Beyond Memorization: Extending Reasoning Depth with Recurrence, Memory and Test-Time Compute Scaling","authors":"","year":null,"url":"https://arxiv.org/abs/2508.16745","region":"t27","hemi":"llm","lobe":"occipital","color":"#6b67b0","indegree":2,"alsoIn":[],"x":0.8234,"y":0.8021,"text":"The paper studies whether and how neural models learn genuine multi-step reasoning rather than memorization, using a controlled 1D cellular-automata (1dCA) benchmark with disjoint train/test rule sets so that solving requires inferring a hidden local Boolean rule and chaining it for k future steps. Most architectures trained from scratch (GPT-NeoX, LSTM, Mamba, ARMT) learn rule inference and achieve near-perfect single-step (k=1) accuracy, but accuracy collapses sharply for k\u22652, revealing a depth barrier; LLMs (except Gemini 2.5 Pro) fail even the simplest radius-1 natural-language 'Handsup' proxy. The authors show model depth\u2014not width\u2014drives multi-step accuracy, and that extending effective depth via recurrence (ARMT), Adaptive Computation Time (ACT), reinforcement learning (GRPO), or token-level Chain-of-Thought pushes the reasoning frontier further (CoT reaching ~k=4), though each remains bounded."},{"id":"0a7193294bbb5d1f9097552f2242f34f","kind":"entity","entity_type":"Topic","n":562,"label":"benchmark","authors":"","year":null,"url":"","region":"t30","hemi":"llm","lobe":"temporal","color":"#38bdf8","indegree":7342,"alsoIn":[],"x":0.5755,"y":0.4405,"text":"benchmark"},{"id":"c5c0a185adae51f49af390e3d5f679c6","kind":"entity","entity_type":"Domain","n":563,"label":"robotics","authors":"","year":null,"url":"","region":"t0","hemi":"wm","lobe":"frontal","color":"#f59e0b","indegree":3789,"alsoIn":[],"x":0.4281,"y":0.5282,"text":"robotics"},{"id":"00ed333d471b517cacd77ef474592c37","kind":"entity","entity_type":"Topic","n":564,"label":"planning","authors":"","year":null,"url":"","region":"t17","hemi":"wm","lobe":"parietal","color":"#38bdf8","indegree":3699,"alsoIn":[],"x":0.351,"y":0.4837,"text":"planning"},{"id":"1ea0320008f054d6a48355261be0862a","kind":"entity","entity_type":"Topic","n":565,"label":"embodied-ai","authors":"","year":null,"url":"","region":"t0","hemi":"wm","lobe":"frontal","color":"#38bdf8","indegree":3523,"alsoIn":[],"x":0.3551,"y":0.4083,"text":"embodied-ai"},{"id":"e2e9917abed652b59f2892d75ccdec63","kind":"entity","entity_type":"Topic","n":566,"label":"llm-agent","authors":"","year":null,"url":"","region":"t19","hemi":"llm","lobe":"occipital","color":"#38bdf8","indegree":3458,"alsoIn":[],"x":0.4058,"y":0.4041,"text":"llm-agent"},{"id":"ef9d9f01afc951daaba0da0ceaeb4988","kind":"entity","entity_type":"Topic","n":567,"label":"dataset","authors":"","year":null,"url":"","region":"t30","hemi":"llm","lobe":"temporal","color":"#38bdf8","indegree":3438,"alsoIn":[],"x":0.4835,"y":0.4515,"text":"dataset"},{"id":"5e88711b4de2546296d173f144a8ca38","kind":"entity","entity_type":"Topic","n":568,"label":"world-model","authors":"","year":null,"url":"","region":"t17","hemi":"wm","lobe":"parietal","color":"#38bdf8","indegree":3414,"alsoIn":[],"x":0.555,"y":0.4879,"text":"world-model"},{"id":"6aeb0ae745b350e18820468d8ce4df3e","kind":"entity","entity_type":"Topic","n":569,"label":"reinforcement-learning","authors":"","year":null,"url":"","region":"t26","hemi":"wm","lobe":"temporal","color":"#38bdf8","indegree":3137,"alsoIn":[],"x":0.3659,"y":0.3897,"text":"reinforcement-learning"},{"id":"1fbb24e7e94e58e18899bd2b06c8fed4","kind":"entity","entity_type":"Topic","n":570,"label":"physical-reasoning","authors":"","year":null,"url":"","region":"t0","hemi":"wm","lobe":"frontal","color":"#38bdf8","indegree":3088,"alsoIn":[],"x":0.563,"y":0.573,"text":"physical-reasoning"},{"id":"b76fb931c1c75ba7964ec119fc2a0b34","kind":"entity","entity_type":"Topic","n":571,"label":"mllm","authors":"","year":null,"url":"","region":"t30","hemi":"llm","lobe":"temporal","color":"#38bdf8","indegree":2978,"alsoIn":[],"x":0.7066,"y":0.4796,"text":"mllm"},{"id":"578297fe62465056b5d74185e59b4615","kind":"entity","entity_type":"Topic","n":572,"label":"robotics","authors":"","year":null,"url":"","region":"t0","hemi":"wm","lobe":"frontal","color":"#38bdf8","indegree":2665,"alsoIn":[],"x":0.3699,"y":0.3564,"text":"robotics"},{"id":"b091ad18c2395679bfef4cda988e8378","kind":"entity","entity_type":"Topic","n":573,"label":"diffusion","authors":"","year":null,"url":"","region":"t32","hemi":"wm","lobe":"frontal","color":"#38bdf8","indegree":2432,"alsoIn":[],"x":0.3643,"y":0.1627,"text":"diffusion"},{"id":"3309b4841a9456de8a44882428576753","kind":"entity","entity_type":"Topic","n":574,"label":"vlm","authors":"","year":null,"url":"","region":"t30","hemi":"llm","lobe":"temporal","color":"#38bdf8","indegree":2376,"alsoIn":[],"x":0.508,"y":0.3773,"text":"vlm"},{"id":"3e5f643d839e5aa2b8723a46fba6a2a6","kind":"entity","entity_type":"Topic","n":575,"label":"video-generation","authors":"","year":null,"url":"","region":"t29","hemi":"wm","lobe":"parietal","color":"#38bdf8","indegree":2182,"alsoIn":[],"x":0.365,"y":0.3808,"text":"video-generation"},{"id":"cf4ebe8a6fd0551fa06e22cccb6525d9","kind":"entity","entity_type":"Topic","n":576,"label":"memory","authors":"","year":null,"url":"","region":"t28","hemi":"llm","lobe":"frontal","color":"#38bdf8","indegree":2068,"alsoIn":[],"x":0.4834,"y":0.4737,"text":"memory"},{"id":"afbadd921ac75a03a7cb4ed228efc708","kind":"entity","entity_type":"Topic","n":577,"label":"inference-engine","authors":"","year":null,"url":"","region":"t24","hemi":"llm","lobe":"frontal","color":"#38bdf8","indegree":1980,"alsoIn":[],"x":0.6946,"y":0.3207,"text":"inference-engine"},{"id":"5d706c4d8e3b5c51bf1534e0ac3649be","kind":"entity","entity_type":"Topic","n":578,"label":"interpretability","authors":"","year":null,"url":"","region":"t16","hemi":"llm","lobe":"frontal","color":"#38bdf8","indegree":1842,"alsoIn":[],"x":0.5969,"y":0.5507,"text":"interpretability"},{"id":"d613731276125291ae874322d754e3e0","kind":"entity","entity_type":"Domain","n":579,"label":"video-generation","authors":"","year":null,"url":"","region":"t29","hemi":"wm","lobe":"parietal","color":"#f59e0b","indegree":1701,"alsoIn":[],"x":0.4214,"y":0.484,"text":"video-generation"},{"id":"f7d9b5d7ab0e51f8a085ffe40543f902","kind":"entity","entity_type":"Topic","n":580,"label":"3d","authors":"","year":null,"url":"","region":"t15","hemi":"wm","lobe":"occipital","color":"#38bdf8","indegree":1648,"alsoIn":[],"x":0.5014,"y":0.2912,"text":"3d"},{"id":"d5699c11ff3e5f2ea251177377b18fff","kind":"entity","entity_type":"Domain","n":581,"label":"embodied-ai","authors":"","year":null,"url":"","region":"t0","hemi":"wm","lobe":"frontal","color":"#f59e0b","indegree":1500,"alsoIn":[],"x":0.4787,"y":0.5188,"text":"embodied-ai"},{"id":"e0312e04abde5e53a1f0b3a14d1422ba","kind":"entity","entity_type":"Topic","n":582,"label":"safety","authors":"","year":null,"url":"","region":"t8","hemi":"llm","lobe":"frontal","color":"#38bdf8","indegree":1308,"alsoIn":[],"x":0.4357,"y":0.3211,"text":"safety"},{"id":"0bf069239279577ebdb04c2e7f09c1c9","kind":"entity","entity_type":"Domain","n":583,"label":"autonomous-driving","authors":"","year":null,"url":"","region":"t17","hemi":"wm","lobe":"parietal","color":"#f59e0b","indegree":1296,"alsoIn":[],"x":0.3319,"y":0.3078,"text":"autonomous-driving"},{"id":"5a9f7bc5b9f45daca6a11a434b3f61d9","kind":"entity","entity_type":"Topic","n":584,"label":"neuroscience","authors":"","year":null,"url":"","region":"t28","hemi":"llm","lobe":"frontal","color":"#38bdf8","indegree":1294,"alsoIn":[],"x":0.6058,"y":0.5827,"text":"neuroscience"},{"id":"2fe319dfa7aa5c34b3373b94d57bfacc","kind":"entity","entity_type":"Domain","n":585,"label":"games","authors":"","year":null,"url":"","region":"t26","hemi":"wm","lobe":"temporal","color":"#f59e0b","indegree":1271,"alsoIn":[],"x":0.2986,"y":0.3854,"text":"games"},{"id":"179eca3284ca53d1909a4865bb5c5f67","kind":"entity","entity_type":"Topic","n":586,"label":"vla","authors":"","year":null,"url":"","region":"t0","hemi":"wm","lobe":"frontal","color":"#38bdf8","indegree":1039,"alsoIn":[],"x":0.3664,"y":0.6191,"text":"vla"},{"id":"41a9b2d7e67f56009ef5f9b25218ae23","kind":"entity","entity_type":"Topic","n":587,"label":"survey","authors":"","year":null,"url":"","region":"t5","hemi":"llm","lobe":"parietal","color":"#38bdf8","indegree":890,"alsoIn":[],"x":0.6289,"y":0.6366,"text":"survey"},{"id":"55dc64e751fc5cada055cbef1ba6f490","kind":"entity","entity_type":"Topic","n":588,"label":"harness","authors":"","year":null,"url":"","region":"t35","hemi":"llm","lobe":"occipital","color":"#38bdf8","indegree":785,"alsoIn":[],"x":0.7207,"y":0.853,"text":"harness"},{"id":"156bd9a2ef6c50b78065af31965a80e8","kind":"entity","entity_type":"Topic","n":589,"label":"autonomous-driving","authors":"","year":null,"url":"","region":"t17","hemi":"wm","lobe":"parietal","color":"#38bdf8","indegree":777,"alsoIn":[],"x":0.3982,"y":0.3349,"text":"autonomous-driving"},{"id":"20e90bf1db27548698a1a21ebe2a12da","kind":"entity","entity_type":"Domain","n":590,"label":"neuroscience","authors":"","year":null,"url":"","region":"t28","hemi":"llm","lobe":"frontal","color":"#f59e0b","indegree":568,"alsoIn":[],"x":0.6002,"y":0.5274,"text":"neuroscience"},{"id":"dfdc4dcc97d55b138155289efc1f8f15","kind":"entity","entity_type":"Domain","n":591,"label":"image-generation","authors":"","year":null,"url":"","region":"t32","hemi":"wm","lobe":"frontal","color":"#f59e0b","indegree":498,"alsoIn":[],"x":0.457,"y":0.2217,"text":"image-generation"},{"id":"6efe5430e683589693361dbeaf1aa256","kind":"entity","entity_type":"Domain","n":592,"label":"natural-language-processing","authors":"","year":null,"url":"","region":"t16","hemi":"llm","lobe":"frontal","color":"#f59e0b","indegree":473,"alsoIn":[],"x":0.785,"y":0.4811,"text":"natural-language-processing"},{"id":"3abafab4d0695bef9da5342372fe06ce","kind":"entity","entity_type":"Domain","n":593,"label":"code-generation","authors":"","year":null,"url":"","region":"t1","hemi":"llm","lobe":"parietal","color":"#f59e0b","indegree":448,"alsoIn":[],"x":0.7751,"y":0.6813,"text":"code-generation"},{"id":"da6911ef37435c6abb169e201aa5f45d","kind":"entity","entity_type":"Domain","n":594,"label":"reinforcement-learning","authors":"","year":null,"url":"","region":"t26","hemi":"wm","lobe":"temporal","color":"#f59e0b","indegree":436,"alsoIn":[],"x":0.3091,"y":0.3784,"text":"reinforcement-learning"},{"id":"6f9da474895254f89a0b8c2c1d528480","kind":"entity","entity_type":"Domain","n":595,"label":"medical","authors":"","year":null,"url":"","region":"t13","hemi":"llm","lobe":"parietal","color":"#f59e0b","indegree":419,"alsoIn":[],"x":0.6985,"y":0.5474,"text":"medical"},{"id":"1320c285fbd7589ea3b5acd47bdc24e5","kind":"entity","entity_type":"Domain","n":596,"label":"question-answering","authors":"","year":null,"url":"","region":"t31","hemi":"llm","lobe":"occipital","color":"#f59e0b","indegree":379,"alsoIn":[],"x":0.7823,"y":0.5156,"text":"question-answering"},{"id":"a4a21ac4385e5f19aae2c6647ebddd70","kind":"entity","entity_type":"Topic","n":597,"label":"theory","authors":"","year":null,"url":"","region":"t26","hemi":"wm","lobe":"temporal","color":"#38bdf8","indegree":376,"alsoIn":[],"x":0.6349,"y":0.5812,"text":"theory"},{"id":"de68e0ca85d358759a695e4fe668563e","kind":"entity","entity_type":"Domain","n":598,"label":"visual-question-answering","authors":"","year":null,"url":"","region":"t30","hemi":"llm","lobe":"temporal","color":"#f59e0b","indegree":366,"alsoIn":[],"x":0.7717,"y":0.2182,"text":"visual-question-answering"},{"id":"07c2931e095759c7a2114e8d9c086a6b","kind":"entity","entity_type":"Domain","n":599,"label":"speech","authors":"","year":null,"url":"","region":"t11","hemi":"llm","lobe":"occipital","color":"#f59e0b","indegree":362,"alsoIn":[],"x":0.6785,"y":0.3993,"text":"speech"},{"id":"3a73876784785037a96bc25790a0f5f4","kind":"entity","entity_type":"Domain","n":600,"label":"video-understanding","authors":"","year":null,"url":"","region":"t29","hemi":"wm","lobe":"parietal","color":"#f59e0b","indegree":356,"alsoIn":[],"x":0.5346,"y":0.4173,"text":"video-understanding"},{"id":"a7af1de80dc55255a1b68d98e0e459d0","kind":"entity","entity_type":"Topic","n":601,"label":"4d","authors":"","year":null,"url":"","region":"t15","hemi":"wm","lobe":"occipital","color":"#38bdf8","indegree":341,"alsoIn":[],"x":0.3621,"y":0.5385,"text":"4d"},{"id":"e4011f27b2e150bea808160e79f649c6","kind":"entity","entity_type":"Domain","n":602,"label":"image-classification","authors":"","year":null,"url":"","region":"t23","hemi":"wm","lobe":"occipital","color":"#f59e0b","indegree":313,"alsoIn":[],"x":0.465,"y":0.366,"text":"image-classification"},{"id":"25f5883a505b5c90b3288786468e427b","kind":"entity","entity_type":"Domain","n":603,"label":"mathematical-reasoning","authors":"","year":null,"url":"","region":"t27","hemi":"llm","lobe":"occipital","color":"#f59e0b","indegree":293,"alsoIn":[],"x":0.7844,"y":0.7144,"text":"mathematical-reasoning"},{"id":"11e409aa7db05ac6926994cc5efd7a4e","kind":"entity","entity_type":"Domain","n":604,"label":"computer-vision","authors":"","year":null,"url":"","region":"t23","hemi":"wm","lobe":"occipital","color":"#f59e0b","indegree":290,"alsoIn":[],"x":0.6588,"y":0.5466,"text":"computer-vision"},{"id":"8b61b4feb56658e4877bb944c345a611","kind":"entity","entity_type":"Domain","n":605,"label":"robotic-manipulation","authors":"","year":null,"url":"","region":"t0","hemi":"wm","lobe":"frontal","color":"#f59e0b","indegree":274,"alsoIn":[],"x":0.2204,"y":0.6681,"text":"robotic-manipulation"},{"id":"5a237ad74e26549290d5b57774b015da","kind":"entity","entity_type":"Domain","n":606,"label":"llm-serving","authors":"","year":null,"url":"","region":"t24","hemi":"llm","lobe":"frontal","color":"#f59e0b","indegree":257,"alsoIn":[],"x":0.8919,"y":0.5163,"text":"llm-serving"},{"id":"d8c361b6739f56a582950c29041b3153","kind":"entity","entity_type":"Domain","n":607,"label":"healthcare","authors":"","year":null,"url":"","region":"t13","hemi":"llm","lobe":"parietal","color":"#f59e0b","indegree":255,"alsoIn":[],"x":0.5328,"y":0.3388,"text":"healthcare"},{"id":"29e518447cab5ffeab46a89a0e892ece","kind":"entity","entity_type":"Domain","n":608,"label":"llm-inference","authors":"","year":null,"url":"","region":"t9","hemi":"llm","lobe":"parietal","color":"#f59e0b","indegree":252,"alsoIn":[],"x":0.8604,"y":0.6031,"text":"llm-inference"},{"id":"6cb325eebf8b5cf1b32fd77c46351199","kind":"entity","entity_type":"Domain","n":609,"label":"software-engineering","authors":"","year":null,"url":"","region":"t35","hemi":"llm","lobe":"occipital","color":"#f59e0b","indegree":242,"alsoIn":[],"x":0.8176,"y":0.7358,"text":"software-engineering"},{"id":"33ec7dab157f55ae97bffe3facf6a915","kind":"entity","entity_type":"Domain","n":610,"label":"education","authors":"","year":null,"url":"","region":"t33","hemi":"llm","lobe":"parietal","color":"#f59e0b","indegree":230,"alsoIn":[],"x":0.7551,"y":0.6834,"text":"education"},{"id":"acab376f7b675f42a7378c3c877df082","kind":"entity","entity_type":"Domain","n":611,"label":"language-modeling","authors":"","year":null,"url":"","region":"t16","hemi":"llm","lobe":"frontal","color":"#f59e0b","indegree":211,"alsoIn":[],"x":0.7135,"y":0.5135,"text":"language-modeling"},{"id":"ad7614ff9d335520b4e20c7315f706c8","kind":"entity","entity_type":"Domain","n":612,"label":"simulation","authors":"","year":null,"url":"","region":"t15","hemi":"wm","lobe":"occipital","color":"#f59e0b","indegree":210,"alsoIn":[],"x":0.4381,"y":0.7585,"text":"simulation"},{"id":"c7b76e0093985d3f8b914a924c3add9c","kind":"entity","entity_type":"Domain","n":613,"label":"robot-manipulation","authors":"","year":null,"url":"","region":"t0","hemi":"wm","lobe":"frontal","color":"#f59e0b","indegree":209,"alsoIn":[],"x":0.2819,"y":0.5915,"text":"robot-manipulation"},{"id":"95bbe72473cf515aa6679ad38582b671","kind":"entity","entity_type":"Domain","n":614,"label":"ai-safety","authors":"","year":null,"url":"","region":"t8","hemi":"llm","lobe":"frontal","color":"#f59e0b","indegree":202,"alsoIn":[],"x":0.6105,"y":0.5909,"text":"ai-safety"},{"id":"362ca2f019995b339e20389e00590b60","kind":"entity","entity_type":"Domain","n":615,"label":"llm-agents","authors":"","year":null,"url":"","region":"t19","hemi":"llm","lobe":"occipital","color":"#f59e0b","indegree":201,"alsoIn":[],"x":0.6623,"y":0.9131,"text":"llm-agents"},{"id":"c88a7a9d1b07512998b157da626ea698","kind":"entity","entity_type":"Domain","n":616,"label":"vision-language","authors":"","year":null,"url":"","region":"t30","hemi":"llm","lobe":"temporal","color":"#f59e0b","indegree":166,"alsoIn":[],"x":0.4662,"y":0.3257,"text":"vision-language"},{"id":"67bf43b7f5555ac8a73a70ffcc0c7ffc","kind":"entity","entity_type":"Domain","n":617,"label":"conversational-ai","authors":"","year":null,"url":"","region":"t1","hemi":"llm","lobe":"parietal","color":"#f59e0b","indegree":160,"alsoIn":[],"x":0.7428,"y":0.7036,"text":"conversational-ai"},{"id":"844313ebbc105d1b83be279baab0e333","kind":"entity","entity_type":"Domain","n":618,"label":"nlp","authors":"","year":null,"url":"","region":"t16","hemi":"llm","lobe":"frontal","color":"#f59e0b","indegree":159,"alsoIn":[],"x":0.8284,"y":0.7251,"text":"nlp"},{"id":"e017307e8bb95c4cb42008a204189377","kind":"entity","entity_type":"Domain","n":619,"label":"finance","authors":"","year":null,"url":"","region":"t19","hemi":"llm","lobe":"occipital","color":"#f59e0b","indegree":159,"alsoIn":[],"x":0.7573,"y":0.7521,"text":"finance"},{"id":"0cf269af8e0050349e9fa97c24b879f8","kind":"entity","entity_type":"Domain","n":620,"label":"navigation","authors":"","year":null,"url":"","region":"t12","hemi":"wm","lobe":"frontal","color":"#f59e0b","indegree":157,"alsoIn":[],"x":0.1925,"y":0.5062,"text":"navigation"},{"id":"cda7dd561a0953c89d239b7026e2b39d","kind":"entity","entity_type":"Domain","n":621,"label":"continuous-control","authors":"","year":null,"url":"","region":"t26","hemi":"wm","lobe":"temporal","color":"#f59e0b","indegree":154,"alsoIn":[],"x":0.342,"y":0.5179,"text":"continuous-control"},{"id":"56556af84b9d5c8382bad825092997e0","kind":"entity","entity_type":"Domain","n":622,"label":"image-captioning","authors":"","year":null,"url":"","region":"t30","hemi":"llm","lobe":"temporal","color":"#f59e0b","indegree":151,"alsoIn":[],"x":0.7503,"y":0.3301,"text":"image-captioning"},{"id":"f4c0d2e5117b5844aa11a1dd0e8b79c5","kind":"entity","entity_type":"Domain","n":623,"label":"computational-neuroscience","authors":"","year":null,"url":"","region":"t28","hemi":"llm","lobe":"frontal","color":"#f59e0b","indegree":140,"alsoIn":[],"x":0.6163,"y":0.5741,"text":"computational-neuroscience"},{"id":"463f8662c60d5656b5416579617882ef","kind":"entity","entity_type":"Domain","n":624,"label":"information-retrieval","authors":"","year":null,"url":"","region":"t31","hemi":"llm","lobe":"occipital","color":"#f59e0b","indegree":138,"alsoIn":[],"x":0.8611,"y":0.6468,"text":"information-retrieval"},{"id":"3f72bc264cb352c986394fe73e6e5d14","kind":"entity","entity_type":"Domain","n":625,"label":"math-reasoning","authors":"","year":null,"url":"","region":"t27","hemi":"llm","lobe":"occipital","color":"#f59e0b","indegree":137,"alsoIn":[],"x":0.8116,"y":0.8104,"text":"math-reasoning"},{"id":"ebf143af89b85b7baba9b1c7f66524e9","kind":"entity","entity_type":"Domain","n":626,"label":"image-editing","authors":"","year":null,"url":"","region":"t32","hemi":"wm","lobe":"frontal","color":"#f59e0b","indegree":135,"alsoIn":[],"x":0.7631,"y":0.2334,"text":"image-editing"},{"id":"a0a77dd812b6582d9a29b5dd65c8ace3","kind":"entity","entity_type":"Topic","n":627,"label":"self-supervised-learning","authors":"","year":null,"url":"","region":"t23","hemi":"wm","lobe":"occipital","color":"#38bdf8","indegree":124,"alsoIn":[],"x":0.3535,"y":0.4338,"text":"self-supervised-learning"},{"id":"53b1c61910fe500e9bce24ac49a31360","kind":"entity","entity_type":"Topic","n":628,"label":"llm","authors":"","year":null,"url":"","region":"t16","hemi":"llm","lobe":"frontal","color":"#38bdf8","indegree":102,"alsoIn":[],"x":0.7961,"y":0.5806,"text":"llm"},{"id":"b8667390705e55ac9928e8cf9a645132","kind":"entity","entity_type":"Topic","n":629,"label":"representation-learning","authors":"","year":null,"url":"","region":"t23","hemi":"wm","lobe":"occipital","color":"#38bdf8","indegree":92,"alsoIn":[],"x":0.4507,"y":0.4425,"text":"representation-learning"},{"id":"c3ad4df1e789548683b4021361c66d6b","kind":"entity","entity_type":"Topic","n":630,"label":"image-generation","authors":"","year":null,"url":"","region":"t32","hemi":"wm","lobe":"frontal","color":"#38bdf8","indegree":73,"alsoIn":[],"x":0.5574,"y":0.1975,"text":"image-generation"},{"id":"60faede8ef095b4f9e15856e8dec4bec","kind":"entity","entity_type":"Topic","n":631,"label":"imitation-learning","authors":"","year":null,"url":"","region":"t0","hemi":"wm","lobe":"frontal","color":"#38bdf8","indegree":68,"alsoIn":[],"x":0.2389,"y":0.6289,"text":"imitation-learning"},{"id":"f4fc68cec2a6556792d727805d687e89","kind":"entity","entity_type":"Topic","n":632,"label":"mixture-of-experts","authors":"","year":null,"url":"","region":"t16","hemi":"llm","lobe":"frontal","color":"#38bdf8","indegree":58,"alsoIn":[],"x":0.7378,"y":0.4525,"text":"mixture-of-experts"},{"id":"b8c5abc6079955949e89e8e5737f65cc","kind":"entity","entity_type":"Topic","n":633,"label":"quantization","authors":"","year":null,"url":"","region":"t37","hemi":"llm","lobe":"parietal","color":"#38bdf8","indegree":54,"alsoIn":[],"x":0.8794,"y":0.3824,"text":"quantization"},{"id":"dfbfd28709bd5ca4b61587004947d0fc","kind":"entity","entity_type":"Topic","n":634,"label":"active-inference","authors":"","year":null,"url":"","region":"t25","hemi":"llm","lobe":"parietal","color":"#38bdf8","indegree":51,"alsoIn":[],"x":0.5834,"y":0.8375,"text":"active-inference"},{"id":"f723aad37a3d5add8b6b8c2871979688","kind":"entity","entity_type":"Topic","n":635,"label":"graph-neural-network","authors":"","year":null,"url":"","region":"t20","hemi":"llm","lobe":"frontal","color":"#38bdf8","indegree":51,"alsoIn":[],"x":0.6121,"y":0.353,"text":"graph-neural-network"},{"id":"43c111a7b6d85e62b33bf10e9ab3cd86","kind":"entity","entity_type":"Topic","n":636,"label":"medical-imaging","authors":"","year":null,"url":"","region":"t13","hemi":"llm","lobe":"parietal","color":"#38bdf8","indegree":50,"alsoIn":[],"x":0.7031,"y":0.1665,"text":"medical-imaging"},{"id":"5aa08fff1eb55ba9ad84ee93a2cc7d04","kind":"entity","entity_type":"Topic","n":637,"label":"speech","authors":"","year":null,"url":"","region":"t11","hemi":"llm","lobe":"occipital","color":"#38bdf8","indegree":49,"alsoIn":[],"x":0.8739,"y":0.2674,"text":"speech"},{"id":"a0eb5cd6f4f955ef910072df8960f881","kind":"entity","entity_type":"Topic","n":638,"label":"medical","authors":"","year":null,"url":"","region":"t13","hemi":"llm","lobe":"parietal","color":"#38bdf8","indegree":49,"alsoIn":[],"x":0.641,"y":0.2008,"text":"medical"},{"id":"dd3ae0bae15e5514884bf73ca72151e5","kind":"entity","entity_type":"Topic","n":639,"label":"code-generation","authors":"","year":null,"url":"","region":"t33","hemi":"llm","lobe":"parietal","color":"#38bdf8","indegree":48,"alsoIn":[],"x":0.8006,"y":0.7931,"text":"code-generation"},{"id":"e520bbf665c753528a6181ff6d7ca2a0","kind":"entity","entity_type":"Topic","n":640,"label":"predictive-coding","authors":"","year":null,"url":"","region":"t20","hemi":"llm","lobe":"frontal","color":"#38bdf8","indegree":48,"alsoIn":[],"x":0.5987,"y":0.6949,"text":"predictive-coding"},{"id":"be97f57d3c57559a855c8ed7dd83f07b","kind":"entity","entity_type":"Topic","n":641,"label":"formal-verification","authors":"","year":null,"url":"","region":"t6","hemi":"llm","lobe":"temporal","color":"#38bdf8","indegree":47,"alsoIn":[],"x":0.814,"y":0.7416,"text":"formal-verification"},{"id":"d6c30301190553349449581c5b70071b","kind":"entity","entity_type":"Topic","n":642,"label":"segmentation","authors":"","year":null,"url":"","region":"t13","hemi":"llm","lobe":"parietal","color":"#38bdf8","indegree":47,"alsoIn":[],"x":0.6346,"y":0.125,"text":"segmentation"},{"id":"576e8eab29425f3db82f2837e1afa170","kind":"entity","entity_type":"Topic","n":643,"label":"contrastive-learning","authors":"","year":null,"url":"","region":"t23","hemi":"wm","lobe":"occipital","color":"#38bdf8","indegree":46,"alsoIn":[],"x":0.5791,"y":0.2647,"text":"contrastive-learning"},{"id":"d3d1df5c9d765256b66104fd1ad76d1d","kind":"entity","entity_type":"Topic","n":644,"label":"healthcare","authors":"","year":null,"url":"","region":"t13","hemi":"llm","lobe":"parietal","color":"#38bdf8","indegree":45,"alsoIn":[],"x":0.7308,"y":0.52,"text":"healthcare"},{"id":"f4282cf7e07b533da381c9a327992940","kind":"entity","entity_type":"Method","n":645,"label":"MotionCtrl","authors":"","year":null,"url":"","region":"t17","hemi":"wm","lobe":"parietal","color":"#34d399","indegree":44,"alsoIn":[],"x":0.6375,"y":0.4181,"text":"MotionCtrl"},{"id":"2547d0055bdf590fbf512b575f47493e","kind":"entity","entity_type":"Topic","n":646,"label":"reasoning","authors":"","year":null,"url":"","region":"t27","hemi":"llm","lobe":"occipital","color":"#38bdf8","indegree":43,"alsoIn":[],"x":0.7273,"y":0.6123,"text":"reasoning"},{"id":"e2cd7f846bb65742a94c810071a398ef","kind":"entity","entity_type":"Topic","n":647,"label":"retrieval-augmented-generation","authors":"","year":null,"url":"","region":"t31","hemi":"llm","lobe":"occipital","color":"#38bdf8","indegree":43,"alsoIn":[],"x":0.8679,"y":0.6064,"text":"retrieval-augmented-generation"},{"id":"12dc907108be5ba6bf916312e91379d8","kind":"entity","entity_type":"Topic","n":648,"label":"multi-agent","authors":"","year":null,"url":"","region":"t22","hemi":"wm","lobe":"temporal","color":"#38bdf8","indegree":42,"alsoIn":[],"x":0.3453,"y":0.469,"text":"multi-agent"},{"id":"94807cc518245d45b6d9b722fdc49931","kind":"entity","entity_type":"Topic","n":649,"label":"multi-agent-systems","authors":"","year":null,"url":"","region":"t22","hemi":"wm","lobe":"temporal","color":"#38bdf8","indegree":40,"alsoIn":[],"x":0.272,"y":0.5188,"text":"multi-agent-systems"},{"id":"8c19be258e0659bfa040d9bd16b2b852","kind":"entity","entity_type":"Topic","n":650,"label":"games","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#38bdf8","indegree":39,"alsoIn":[],"x":0.2645,"y":0.5019,"text":"games"},{"id":"ce96547a8bd75f2c950f698f9e0b42d7","kind":"entity","entity_type":"Method","n":651,"label":"CameraCtrl","authors":"","year":null,"url":"","region":"t17","hemi":"wm","lobe":"parietal","color":"#34d399","indegree":39,"alsoIn":[],"x":0.6375,"y":0.4181,"text":"CameraCtrl"},{"id":"f4da11522c2a5573a1d43f2f81e7df3a","kind":"entity","entity_type":"Topic","n":652,"label":"video-understanding","authors":"","year":null,"url":"","region":"t29","hemi":"wm","lobe":"parietal","color":"#38bdf8","indegree":39,"alsoIn":[],"x":0.272,"y":0.5188,"text":"video-understanding"},{"id":"f5d42622e0125527b372433a7ec6a846","kind":"entity","entity_type":"Topic","n":653,"label":"generative-model","authors":"","year":null,"url":"","region":"t11","hemi":"llm","lobe":"occipital","color":"#38bdf8","indegree":39,"alsoIn":[],"x":0.648,"y":0.5663,"text":"generative-model"},{"id":"e9dea714730e5a2dae3c88f69887a31c","kind":"entity","entity_type":"Method","n":654,"label":"Simulator","authors":"","year":null,"url":"","region":"t17","hemi":"wm","lobe":"parietal","color":"#34d399","indegree":38,"alsoIn":[],"x":0.6375,"y":0.4181,"text":"Simulator"},{"id":"ee22adec46ca5d95a110b38805bb1097","kind":"entity","entity_type":"Method","n":655,"label":"Evaluation engine","authors":"","year":null,"url":"","region":"t17","hemi":"wm","lobe":"parietal","color":"#34d399","indegree":37,"alsoIn":[],"x":0.6375,"y":0.4181,"text":"Evaluation engine"},{"id":"dc86e7812f445fa6a43e21a83dc1b287","kind":"entity","entity_type":"Method","n":656,"label":"Autonomous policy rollouts","authors":"","year":null,"url":"","region":"t17","hemi":"wm","lobe":"parietal","color":"#34d399","indegree":34,"alsoIn":[],"x":0.6375,"y":0.4181,"text":"Autonomous policy rollouts"},{"id":"d49321f4b3105e6f9f83a04623bad056","kind":"entity","entity_type":"Method","n":657,"label":"Backbone","authors":"","year":null,"url":"","region":"t17","hemi":"wm","lobe":"parietal","color":"#34d399","indegree":33,"alsoIn":[],"x":0.6375,"y":0.4181,"text":"Backbone"},{"id":"c6594b9752b051f58bc98e3c76a46cad","kind":"entity","entity_type":"Method","n":658,"label":"production-level evaluation","authors":"","year":null,"url":"","region":"t17","hemi":"wm","lobe":"parietal","color":"#34d399","indegree":32,"alsoIn":[],"x":0.6375,"y":0.4181,"text":"production-level evaluation"},{"id":"cd4ebce009bc570da4a44f22bbdf9ba8","kind":"entity","entity_type":"Method","n":659,"label":"ManiSkill","authors":"","year":null,"url":"","region":"t17","hemi":"wm","lobe":"parietal","color":"#34d399","indegree":31,"alsoIn":[],"x":0.6375,"y":0.4181,"text":"ManiSkill"},{"id":"0c7c7b4eb4b5569dbff6d810e7784c49","kind":"entity","entity_type":"Method","n":660,"label":"DINOv2","authors":"","year":null,"url":"","region":"t23","hemi":"wm","lobe":"occipital","color":"#34d399","indegree":28,"alsoIn":[],"x":0.2407,"y":0.3462,"text":"DINOv2"},{"id":"5ab80843a1a1566ebb7cd7bdb5120e5d","kind":"entity","entity_type":"Method","n":661,"label":"Training Compute-Optimal Large Language Models","authors":"","year":null,"url":"","region":"t17","hemi":"wm","lobe":"parietal","color":"#34d399","indegree":23,"alsoIn":[],"x":0.6375,"y":0.4181,"text":"Training Compute-Optimal Large Language Models"},{"id":"35a196c11652577b9801ae4fefc157d0","kind":"entity","entity_type":"Method","n":662,"label":"Orbit","authors":"","year":null,"url":"","region":"t17","hemi":"wm","lobe":"parietal","color":"#34d399","indegree":16,"alsoIn":[],"x":0.6375,"y":0.4181,"text":"Orbit"},{"id":"4068f7c4e1f6531daba0ce4f4275a89b","kind":"entity","entity_type":"Term","n":663,"label":"Marble","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":15,"alsoIn":[],"x":0.4208,"y":0.9271,"text":"World Labs' generative multimodal world model that creates, edits, expands, composes, and exports explorable 3D worlds from text, image, video, or coarse 3D layout inputs."},{"id":"637df3c30efd5f9f91bc799b491146f1","kind":"entity","entity_type":"Method","n":664,"label":"Reinforcement learning","authors":"","year":null,"url":"","region":"t19","hemi":"llm","lobe":"occipital","color":"#34d399","indegree":14,"alsoIn":[],"x":0.6653,"y":0.8372,"text":"Reinforcement learning"},{"id":"c9df46f9eaf15c4d9c3a5ce01a64c1b7","kind":"entity","entity_type":"Method","n":665,"label":"GENSIM","authors":"","year":null,"url":"","region":"t0","hemi":"wm","lobe":"frontal","color":"#34d399","indegree":14,"alsoIn":[],"x":0.1745,"y":0.6867,"text":"GENSIM"},{"id":"6318ed26b8415943b81b6e73282339cc","kind":"entity","entity_type":"Method","n":666,"label":"Generative Agents","authors":"","year":null,"url":"","region":"t19","hemi":"llm","lobe":"occipital","color":"#34d399","indegree":13,"alsoIn":[],"x":0.6221,"y":0.8834,"text":"Generative Agents"},{"id":"486baa7031875dadadd3ee7d83a7a94d","kind":"entity","entity_type":"Term","n":667,"label":"JEPA (Joint Embedding Predictive Architecture)","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":11,"alsoIn":[],"x":0.6742,"y":0.6077,"text":"LeCun/AMI Labs approach that learns abstract representations of physical reality in a compressed embedding space rather than predicting outputs token by token."},{"id":"e64a473ca88450d0bd31bf1a8b7118e3","kind":"entity","entity_type":"Method","n":668,"label":"Stochastic Gradient Variational Bayes estimator","authors":"","year":null,"url":"","region":"t3","hemi":"wm","lobe":"occipital","color":"#34d399","indegree":11,"alsoIn":[],"x":0.648,"y":0.5663,"text":"Stochastic Gradient Variational Bayes estimator"},{"id":"ef539d7d7d345e4d816f27b3a4120bcb","kind":"entity","entity_type":"Method","n":669,"label":"Curriculum learning","authors":"","year":null,"url":"","region":"t26","hemi":"wm","lobe":"temporal","color":"#34d399","indegree":11,"alsoIn":[],"x":0.502,"y":0.5849,"text":"Curriculum learning"},{"id":"24c813c93c21563b8f444f5c8a02ad89","kind":"entity","entity_type":"Method","n":670,"label":"Energy-Based Models","authors":"","year":null,"url":"","region":"t14","hemi":"llm","lobe":"temporal","color":"#34d399","indegree":10,"alsoIn":[],"x":0.6705,"y":0.5734,"text":"Energy-Based Models"},{"id":"77064b78bc7458b28cf78b7f3e5eeb73","kind":"entity","entity_type":"Method","n":671,"label":"AdaPlanner","authors":"","year":null,"url":"","region":"t19","hemi":"llm","lobe":"occipital","color":"#34d399","indegree":10,"alsoIn":[],"x":0.5623,"y":0.9412,"text":"AdaPlanner"},{"id":"a65e041dcf2e5692a4e593140113a16d","kind":"entity","entity_type":"Method","n":672,"label":"Denoising Autoencoders","authors":"","year":null,"url":"","region":"t23","hemi":"wm","lobe":"occipital","color":"#34d399","indegree":10,"alsoIn":[],"x":0.4042,"y":0.2076,"text":"Denoising Autoencoders"},{"id":"cc066218775353349cac77b5cf3091a3","kind":"entity","entity_type":"Method","n":673,"label":"Pre-Training","authors":"","year":null,"url":"","region":"t1","hemi":"llm","lobe":"parietal","color":"#34d399","indegree":10,"alsoIn":[],"x":0.8177,"y":0.5876,"text":"Pre-Training"},{"id":"817a3ffd1c285c55b8a2eb938ba85f4e","kind":"entity","entity_type":"Term","n":674,"label":"Oasis","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":9,"alsoIn":[],"x":0.3823,"y":0.705,"text":"Decart model line built on DOS for physical AI (robotics, autonomous vehicles) via physically accurate real-time simulation"},{"id":"c2c0e7d5306a5897b8bc22c8fce99645","kind":"entity","entity_type":"Term","n":675,"label":"Genie 3","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":9,"alsoIn":[],"x":0.3326,"y":0.6696,"text":"Google DeepMind's transformer-based generative world model that creates navigable 3D worlds frame-by-frame at 24 fps in 720p with promptable world events."},{"id":"d8e9b19f33135d92b63f857a31154224","kind":"entity","entity_type":"Term","n":676,"label":"Action Expert","authors":"","year":null,"url":"","region":"t0","hemi":"wm","lobe":"frontal","color":"#e879f9","indegree":8,"alsoIn":[],"x":0.148,"y":0.7693,"text":"A flow-matching-based module that augments the pretrained VLM with continuous action outputs at up to 50 Hz for dexterous robot control."},{"id":"1106a1b9f3515e9290d700aa9e21f35b","kind":"entity","entity_type":"Term","n":677,"label":"Muse","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":7,"alsoIn":[],"x":0.272,"y":0.5188,"text":"Microsoft Research / Ninja Theory generative AI model for gameplay ideation, trained on Bleeding Edge."},{"id":"5547843881175855b026a2df0af18c7a","kind":"entity","entity_type":"Term","n":678,"label":"Waabi World","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":7,"alsoIn":[],"x":0.272,"y":0.5188,"text":"Waabi's neural controllable simulator for autonomous vehicle testing, validated at 99.7% realism via pair-setting."},{"id":"5e29463a898b5d3ab547102ee9ce6bfd","kind":"entity","entity_type":"Term","n":679,"label":"Cosmos 3","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":7,"alsoIn":[],"x":0.4315,"y":0.9533,"text":"Family of omnimodal world models (Edge 4B, Nano 16B, Super 64B) jointly modeling language, image, video, audio, and action for understanding and generation."},{"id":"addc77b5856b50cd8a6642888e8d8b96","kind":"entity","entity_type":"Term","n":680,"label":"Three-stage training pipeline","authors":"","year":null,"url":"","region":"t30","hemi":"llm","lobe":"temporal","color":"#e879f9","indegree":7,"alsoIn":[],"x":0.7308,"y":0.52,"text":"Stage I learns the frozen low-level tokenizer; Stage II discovers skill latents and pretrains planner/executor without memory; Stage III finetunes the full policy with the gated memory module."},{"id":"b857f1ae5f585f12bed9b48d888ca7ab","kind":"entity","entity_type":"Term","n":681,"label":"V-JEPA 2","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":7,"alsoIn":[],"x":0.2833,"y":0.4774,"text":"Next-generation 1.2B-parameter V-JEPA pretrained on 22M videos, fine-tunable with ~60h of robot data for manipulation planning."},{"id":"cb06a17fae925dbeb50d75fd3daf1b1c","kind":"entity","entity_type":"Term","n":682,"label":"Two-stage training strategy","authors":"","year":null,"url":"","region":"t0","hemi":"wm","lobe":"frontal","color":"#e879f9","indegree":7,"alsoIn":[],"x":0.272,"y":0.5188,"text":"Stage 1 text-to-video pretraining of the base model on OGameData-GEN; Stage 2 freezes the base and trains only InstructNet on OGameData-INS for control."},{"id":"26a42a6a9311514da6bbc37bba8457ec","kind":"entity","entity_type":"Term","n":683,"label":"GWM-1","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":6,"alsoIn":[],"x":0.272,"y":0.5188,"text":"Runway's first General World Model, built on Gen-4.5, that generates video frame by frame with interactive control via camera, robot commands, or audio."},{"id":"8d34e1fc3d105707a21910c27e565b82","kind":"entity","entity_type":"Term","n":684,"label":"PROWL","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":6,"alsoIn":[],"x":0.272,"y":0.5188,"text":"An RL-driven adversarial framework where an RL agent explores game environments to improve world model performance"},{"id":"9df51276f50f503db60d7f16830fbc6b","kind":"entity","entity_type":"Term","n":685,"label":"Waabi Driver","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":6,"alsoIn":[],"x":0.272,"y":0.5188,"text":"Waabi's autonomy system / self-driving software stack that is tested and trained inside Waabi World."},{"id":"c483220c93825f1bab9625e6d8aecd9f","kind":"entity","entity_type":"Term","n":686,"label":"Active inference","authors":"","year":null,"url":"","region":"t25","hemi":"llm","lobe":"parietal","color":"#e879f9","indegree":6,"alsoIn":[],"x":0.5673,"y":0.8404,"text":"Friston's overarching theory unifying predictive processing with action, where agents act on the environment to bring sensations in line with inviolable internal model predictions."},{"id":"ec00eb3c034e5985bdb7ddf4866f493e","kind":"entity","entity_type":"Term","n":687,"label":"Three-stage training strategy","authors":"","year":null,"url":"","region":"t17","hemi":"wm","lobe":"parietal","color":"#e879f9","indegree":6,"alsoIn":[],"x":0.272,"y":0.5188,"text":"Pretraining (~500k) to activate multi-view visual generation, SFT cold-start on multi-task mixed data, and GRPO reinforcement learning to explore reasoning."},{"id":"f11c75d47c2e53a4b94fc0a39810c80e","kind":"entity","entity_type":"Term","n":688,"label":"Genie","authors":"","year":null,"url":"","region":"t32","hemi":"wm","lobe":"frontal","color":"#e879f9","indegree":6,"alsoIn":[],"x":0.3326,"y":0.639,"text":"Google's generative interactive environment trained on internet video data, using ST-Transformer, latent action model (codebook size 8), video tokenizer (codebook size 1024), and MaskGIT-based dynamics model."},{"id":"142d56d7638b53659e24beadb82517fb","kind":"entity","entity_type":"Term","n":689,"label":"Promptable world events","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":5,"alsoIn":[],"x":0.3326,"y":0.6696,"text":"Ability to use a text prompt to inject changes or interventions into the already-generated world."},{"id":"1499a1f59c7c5b938b63a596dce80547","kind":"entity","entity_type":"Term","n":690,"label":"Cascade","authors":"","year":null,"url":"","region":"t8","hemi":"llm","lobe":"frontal","color":"#e879f9","indegree":5,"alsoIn":[],"x":0.7308,"y":0.52,"text":"A module integrating a Risk Window (sum of last W=5 step scores vs threshold \u03b8=1.0), Selective Episodic Memory (at most 2 recalled prior steps), and a two-tier LLM Judge (gpt-4o-mini cheap / gpt-4o advanced) with mutually exclusive routing."},{"id":"167f7f0675155167b1407dbd05df9210","kind":"entity","entity_type":"Term","n":691,"label":"Ray3","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":5,"alsoIn":[],"x":0.272,"y":0.5188,"text":"Luma AI's next-generation cinematic text/image-to-video model with built-in chain-of-thought reasoning, ~2x the size of Ray2."},{"id":"2265ce6d85fa5f579da9f46b6178054a","kind":"entity","entity_type":"Term","n":692,"label":"Two-Stage training","authors":"","year":null,"url":"","region":"t29","hemi":"wm","lobe":"parietal","color":"#e879f9","indegree":5,"alsoIn":[],"x":0.272,"y":0.5188,"text":"First pretrain the WAM on semantic visual and latent-action tokens, then adapt to embodiment-specific robot-action control, decoupling latent-action pretraining from robot-action adaptation."},{"id":"37bd6a5426ca5981bb60f3765730bd51","kind":"entity","entity_type":"Term","n":693,"label":"Starchild-1","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":5,"alsoIn":[],"x":0.272,"y":0.5188,"text":"A planned model learning from richer multimodal interaction rather than visual observation alone"},{"id":"425e1e76829350378eb6218d132f8d01","kind":"entity","entity_type":"Term","n":694,"label":"Executor","authors":"","year":null,"url":"","region":"t19","hemi":"llm","lobe":"occipital","color":"#e879f9","indegree":5,"alsoIn":[],"x":0.7308,"y":0.52,"text":"Trainable student video model conditioned only on the initial image and a short task instruction, learning to map instructions directly to action trajectories."},{"id":"5b9fd3728a5a5cc6a157247791175001","kind":"entity","entity_type":"Term","n":695,"label":"Planner","authors":"","year":null,"url":"","region":"t0","hemi":"wm","lobe":"frontal","color":"#e879f9","indegree":5,"alsoIn":[],"x":0.272,"y":0.5188,"text":"A world model that takes observations plus a goal and outputs actions, closing the perception-action loop (inverse of a renderer)."},{"id":"715d0b7b2c4f56d4a0d8c9be9684c5e6","kind":"entity","entity_type":"Term","n":696,"label":"Large Geospatial Model (LGM)","authors":"","year":null,"url":"","region":"t12","hemi":"wm","lobe":"frontal","color":"#e879f9","indegree":5,"alsoIn":[],"x":0.5732,"y":0.1731,"text":"A geo-specific deep-learning model of the world built from billions of images to maintain a fresh, reliable representation of reality for localization."},{"id":"8669e2f0da50530491a3851cadce7e1a","kind":"entity","entity_type":"Term","n":697,"label":"Chisel","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":5,"alsoIn":[],"x":0.272,"y":0.5188,"text":"Experimental AI-native 3D editing mode that decouples world structure (coarse 3D shapes/imported assets) from style (text prompt) to sculpt detailed worlds."},{"id":"8de0b6d5720051d8bc1d8de5781c3155","kind":"entity","entity_type":"Term","n":698,"label":"Bayesian mechanics","authors":"","year":null,"url":"","region":"t25","hemi":"llm","lobe":"parietal","color":"#e879f9","indegree":5,"alsoIn":[],"x":0.5792,"y":0.8107,"text":"Framework in which the internal states of a Markov-blanketed system at (Gaussian) steady state appear to perform variational Bayesian inference over external states."},{"id":"983982f2750e51e2ac18870a3e1238d7","kind":"entity","entity_type":"Term","n":699,"label":"Cosmos Reason","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":5,"alsoIn":[],"x":0.272,"y":0.5188,"text":"Open, customizable WFM with spatiotemporal awareness using chain-of-thought reasoning to understand video and predict interaction outcomes in natural language."},{"id":"bf0453e0344b54c99629b05d71d88721","kind":"entity","entity_type":"Term","n":700,"label":"Neural Generative Coding (NGC)","authors":"","year":null,"url":"","region":"t20","hemi":"llm","lobe":"frontal","color":"#e879f9","indegree":5,"alsoIn":[],"x":0.5934,"y":0.6112,"text":"A computational framework for learning generative ANNs inspired by predictive processing, using local error neurons and Hebbian-like synaptic updates instead of backpropagation."},{"id":"e9707c46872b54fc8d0ab6898b5a8b66","kind":"entity","entity_type":"Term","n":701,"label":"Agora-1","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":5,"alsoIn":[],"x":0.272,"y":0.5188,"text":"A multi-agent world model letting multiple human or AI participants share and interact in the same real-time world simulation"},{"id":"ef96446cbb435bdc87f8701e55712374","kind":"entity","entity_type":"Term","n":702,"label":"WHAM (World and Human Action Model)","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":5,"alsoIn":[],"x":0.272,"y":0.5188,"text":"Model class that jointly captures the game world dynamics, physics, and human controller actions to generate gameplay."},{"id":"efc5d4f358ce5894b78284ab94cc7bda","kind":"entity","entity_type":"Term","n":703,"label":"data engine","authors":"","year":null,"url":"","region":"t30","hemi":"llm","lobe":"temporal","color":"#e879f9","indegree":5,"alsoIn":[],"x":0.7308,"y":0.52,"text":"A released video-reconstruction pipeline that auto-estimates camera pose and metric depth on arbitrary video (using VGGT, MoGe, Metric3D) to produce training data without manual 3D annotation."},{"id":"f633af356dbf5f378271c0dc2800d089","kind":"entity","entity_type":"Term","n":704,"label":"Odyssey-2","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":5,"alsoIn":[],"x":0.272,"y":0.5188,"text":"Odyssey's next-generation, more powerful general-purpose world model advancing physical accuracy"},{"id":"02e907aa5b48500e87e03a1cd9b4c138","kind":"entity","entity_type":"Term","n":705,"label":"Memory module","authors":"","year":null,"url":"","region":"t19","hemi":"llm","lobe":"occipital","color":"#e879f9","indegree":4,"alsoIn":[],"x":0.7308,"y":0.52,"text":"Persistent multimodal state center storing history, retrieving context, updating state, and managing sessions for multi-turn interaction."},{"id":"02f7fb51dd405cf7a36ebcdf961d23fd","kind":"entity","entity_type":"Term","n":706,"label":"Lucy","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":4,"alsoIn":[],"x":0.272,"y":0.5188,"text":"Decart model line built on DOS for real-time immersive experiences in gaming, e-commerce, and advertising with sub-30ms response times"},{"id":"0bf8bc0d26195a2da57b02e5fc304610","kind":"entity","entity_type":"Term","n":707,"label":"spatial intelligence","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":4,"alsoIn":[],"x":0.272,"y":0.5188,"text":"The ability of machines to process visual data, make predictions, and act upon those predictions in the physical world."},{"id":"0d964edb4353516da96e5aee3d389c8c","kind":"entity","entity_type":"Term","n":708,"label":"Latent Action Model (LAM)","authors":"","year":null,"url":"","region":"t0","hemi":"wm","lobe":"frontal","color":"#e879f9","indegree":4,"alsoIn":[],"x":0.3326,"y":0.639,"text":"VQ-VAE-based encoder-decoder that infers latent actions from consecutive frames using a small codebook (size 8) to encode frame-to-frame changes as controllable actions."},{"id":"0fb2127e562b59e5893fc9889bcb5b15","kind":"entity","entity_type":"Term","n":709,"label":"Unified action space","authors":"","year":null,"url":"","region":"t35","hemi":"llm","lobe":"occipital","color":"#e879f9","indegree":4,"alsoIn":[],"x":0.5822,"y":0.213,"text":"Eleven-plus representative actions (tap, swipe, textentry, terminate, open_app, etc.) in a function-call representation spanning shared and platform-specific operations (Table 5)."},{"id":"135ad2f14b9a517fae601d9d928531c3","kind":"entity","entity_type":"Term","n":710,"label":"GWM Robotics","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":4,"alsoIn":[],"x":0.272,"y":0.5188,"text":"Variant of GWM-1 for producing synthetic training data for robots."},{"id":"175e9a9b50aa586bac9fc5c9eb0aec22","kind":"entity","entity_type":"Term","n":711,"label":"VL-JEPA","authors":"","year":null,"url":"","region":"t38","hemi":"wm","lobe":"temporal","color":"#e879f9","indegree":4,"alsoIn":[],"x":0.272,"y":0.5188,"text":"Vision-Language JEPA \u2014 a variant of JEPA that combines vision and language modalities, reasoning via vector embeddings of visual + caption representations rather than discrete tokens."}],"edges":[{"source":"arxiv:1708.00133","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2603.00479","target":"concept_t13","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2305.10601","target":"concept_t27","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2502.12130","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.26802","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.21219","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2312.03815","target":"concept_t35","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2402.01695","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2311.05348","target":"concept_t30","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2401.12945","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2504.00983","target":"concept_t15","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1910.10683","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2512.04797","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"psychclassics.yorku.ca:d7323f879ff8","target":"concept_t5","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2311.07222","target":"concept_t34","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2605.12090","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2506.18897","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2511.00062","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2601.15533","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.04999","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2304.12995","target":"concept_t11","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2602.22010","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2403.15498","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.05865","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"proceedings.mlr.press:6beb908867f9","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2310.02207","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2603.27918","target":"concept_t8","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2602.15400","target":"concept_t12","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2403.06845","target":"concept_t15","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2507.00917","target":"concept_t7","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2402.15391","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2102.01192","target":"concept_t11","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1705.09156","target":"concept_t25","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1706.03741","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2204.03458","target":"concept_t32","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2507.10535","target":"concept_t33","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2005.14165","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2512.00597","target":"concept_t13","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2210.03629","target":"concept_t27","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2202.09467","target":"concept_t20","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2410.02548","target":"concept_t3","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2307.16125","target":"concept_t33","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.20416","target":"concept_t34","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2602.10717","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2407.08867","target":"concept_t7","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2309.02591","target":"concept_t23","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2503.06800","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.03847","target":"concept_t35","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"openreview.net:c15700430713","target":"concept_t15","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2207.12316","target":"concept_t20","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2303.01469","target":"concept_t32","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"transformer-circuits.pub:e2306116790e","target":"concept_t16","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1908.05656","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2410.18072","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"cambridge.org:57359c90d672","target":"concept_t5","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"nature.com:7719482cd2c4","target":"concept_t28","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"transformer-circuits.pub:c4d9a1f0e2b7","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2308.03688","target":"concept_t35","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2309.17080","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"nature.com:7fc0dd8e4038","target":"concept_t25","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2606.24467","target":"concept_t18","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2604.06882","target":"concept_t39","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1803.10122","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2502.11831","target":"concept_t20","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2406.14798","target":"concept_t34","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1312.6114","target":"concept_t3","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2312.03788","target":"concept_t37","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2407.00118","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1709.06560","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2204.00227","target":"concept_t32","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2305.10626","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2506.09985","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2110.04544","target":"concept_t23","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2401.15884","target":"concept_t31","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2501.09038","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2502.05855","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2310.08560","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2305.15074","target":"concept_t33","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2602.24044","target":"concept_t24","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2503.02143","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2103.03725","target":"concept_t20","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.12966","target":"concept_t9","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2604.16592","target":"concept_t7","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2605.02900","target":"concept_t7","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2505.21996","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2507.05201","target":"concept_t13","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2202.08906","target":"concept_t16","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2308.12284","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2208.01769","target":"concept_t22","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2512.02419","target":"concept_t5","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2603.28489","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2511.12239","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2502.03373","target":"concept_t27","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1011.0063","target":"concept_t36","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2605.17288","target":"concept_t8","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1712.04602","target":"concept_t28","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2504.09590","target":"concept_t24","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2605.29930","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2412.03568","target":"concept_t15","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2206.04114","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2411.17470","target":"concept_t32","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2411.07690","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1907.05600","target":"concept_t3","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.01539","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2601.21998","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2305.17198","target":"concept_t22","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2404.12377","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2210.07729","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2209.00588","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2305.18264","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2605.23972","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2407.15549","target":"concept_t8","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"frontiersin.org:d1ec3c5c2c89","target":"concept_t28","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2506.09171","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1912.01603","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1805.09042","target":"concept_t28","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2310.01361","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2401.13178","target":"concept_t35","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2112.07103","target":"concept_t4","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2406.00877","target":"concept_t20","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"nature.com:e56ca063b728","target":"concept_t22","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2310.08803","target":"concept_t5","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2505.02074","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2401.07851","target":"concept_t9","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2309.17382","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2010.11929","target":"concept_t23","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2506.23068","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2508.13073","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2406.16062","target":"concept_t20","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2212.09748","target":"concept_t32","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2401.02954","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2310.06114","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2206.10498","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2305.17926","target":"concept_t33","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2606.20545","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2405.10480","target":"concept_t16","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2203.03417","target":"concept_t4","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2205.06175","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2411.06559","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2508.06297","target":"concept_t18","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2310.08864","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2405.12399","target":"concept_t32","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2505.01712","target":"concept_t39","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2206.14176","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2603.20397","target":"concept_t18","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2408.14837","target":"concept_t32","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2401.09985","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2503.14492","target":"concept_t15","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2007.08794","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2210.04133","target":"concept_t13","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2506.21876","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2403.08540","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2112.15242","target":"concept_t25","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2201.06387","target":"concept_t25","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2309.00941","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1911.11641","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"philpapers.org:7b684275d20a","target":"concept_t2","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"transformer-circuits.pub:b904370d3db5","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"deepmind.google:39914de6d3c9","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1807.11819","target":"concept_t5","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2502.14819","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"nature.com:f005fae15467","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2309.05767","target":"concept_t11","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2602.03916","target":"concept_t12","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2212.12794","target":"concept_t34","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2309.08600","target":"concept_t16","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2502.21321","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"thesequence.substack.com:0f25bda15c4b","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2312.09257","target":"concept_t5","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"reply.com:5748177416db","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"aclanthology.org:8f7acb76ba1d","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2112.10752","target":"concept_t32","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2509.14252","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1903.00374","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2204.01691","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2404.08471","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2509.12387","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2508.12027","target":"concept_t14","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2208.06726","target":"concept_t36","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2303.08302","target":"concept_t37","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2202.03771","target":"concept_t4","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2301.04104","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.13161","target":"concept_t9","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2401.03804","target":"concept_t10","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2411.06284","target":"concept_t30","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2405.15383","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2506.06725","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2406.09455","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"lumalabs.ai:cf7b765e8f15","target":"concept_t3","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"nature.com:386c3df42560","target":"concept_t34","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2006.04182","target":"concept_t20","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2103.04723","target":"concept_t4","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2305.18153","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2403.00504","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"huggingface.co:9f24e732f5fe","target":"concept_t15","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2305.07017","target":"concept_t23","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"doi.org:0a734c9c22c9","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2406.11944","target":"concept_t16","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2206.11795","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2111.00210","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1709.02341","target":"concept_t14","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2311.16038","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2412.18106","target":"concept_t24","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2306.14066","target":"concept_t34","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2505.09694","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2308.16369","target":"concept_t24","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2411.00769","target":"concept_t15","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2305.05383","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2509.19958","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2605.16547","target":"concept_t39","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2501.03575","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2411.10171","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"nianticlabs.com:05a236f27efe","target":"concept_t12","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2407.06886","target":"concept_t7","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.16732","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1810.04805","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2402.16823","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"nature.com:e8ee4994d788","target":"concept_t5","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1911.08265","target":"concept_t22","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2307.15054","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2006.11239","target":"concept_t32","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2408.03943","target":"concept_t7","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2403.16971","target":"concept_t35","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2504.10612","target":"concept_t3","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2105.11203","target":"concept_t25","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2602.23152","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2506.08967","target":"concept_t11","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"the-decoder.com:97885f462d1f","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2410.13056","target":"concept_t37","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2504.10903","target":"concept_t27","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2107.08241","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2507.06952","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2404.07214","target":"concept_t30","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.07092","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2606.00133","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2102.08363","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2111.09259","target":"concept_t5","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2209.15571","target":"concept_t3","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2402.03161","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2605.06192","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2311.15127","target":"concept_t32","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2506.09849","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2410.24164","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2502.13092","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2505.24631","target":"concept_t36","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2106.00737","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2310.16828","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1801.04819","target":"concept_t7","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2605.16395","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"nature.com:a4d8ac562091","target":"concept_t28","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2507.12440","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"openaipublic.blob.core.windows.net:570a1f9c6e10","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2306.05685","target":"concept_t33","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2407.12036","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2305.14992","target":"concept_t27","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2407.10311","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2411.04983","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2606.06556","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2603.13910","target":"concept_t15","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2509.02547","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2501.02189","target":"concept_t30","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2509.13281","target":"concept_t8","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2511.20340","target":"concept_t9","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"github.com:14ee211b0b5d","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"people.idsia.ch:73d35fac50f9","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2506.21521","target":"concept_t33","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2411.08794","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2502.14739","target":"concept_t33","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2101.07393","target":"concept_t22","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2304.03442","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2412.07077","target":"concept_t23","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2503.15558","target":"concept_t7","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"pubmed.ncbi.nlm.nih.gov:fd66789ead66","target":"concept_t28","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2502.10498","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"nature.com:ad56347f1921","target":"concept_t28","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2405.17398","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2507.21141","target":"concept_t8","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2511.20644","target":"concept_t12","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2310.04276","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2511.08585","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2205.11916","target":"concept_t27","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2605.09131","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2406.18043","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2303.00915","target":"concept_t13","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"philpapers.org:9a9c4dcb81d9","target":"concept_t2","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1701.06970","target":"concept_t36","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2411.02385","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"blogs.nvidia.com:9f8a6e618fbd","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2506.22355","target":"concept_t7","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2402.10196","target":"concept_t8","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2304.14772","target":"concept_t3","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2512.21577","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2603.20224","target":"concept_t24","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2302.01560","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2311.01017","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1904.08149","target":"concept_t14","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2507.02076","target":"concept_t27","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"engraved.blog:b8c16edd622e","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"neurips.cc:9c7e4f09bb35","target":"concept_t35","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"huggingface.co:e9042540da12","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2503.20523","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2605.07234","target":"concept_t18","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1910.08210","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2309.12288","target":"concept_t27","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2412.02252","target":"concept_t18","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2407.01392","target":"concept_t32","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2503.20552","target":"concept_t24","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1706.02275","target":"concept_t22","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"philpapers.org:4f7ed9918446","target":"concept_t2","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2307.13692","target":"concept_t33","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2308.09723","target":"concept_t37","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2509.00579","target":"concept_t18","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2310.03744","target":"concept_t30","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1812.00922","target":"concept_t22","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2305.18654","target":"concept_t27","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2301.08243","target":"concept_t23","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"philpapers.org:8c0576aa7f12","target":"concept_t2","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2406.03689","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2405.03520","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"introl.com:c695c416f507","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2603.13647","target":"concept_t39","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2410.05229","target":"concept_t27","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2405.05925","target":"concept_t34","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2405.05967","target":"concept_t32","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2411.14499","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2304.09116","target":"concept_t11","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2403.04253","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2606.18250","target":"concept_t15","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"pubmed.ncbi.nlm.nih.gov:9eeb70dd85f3","target":"concept_t5","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2303.11366","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2503.16416","target":"concept_t35","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2108.13343","target":"concept_t25","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2212.06817","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2505.24618","target":"concept_t39","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1510.00056","target":"concept_t36","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2506.01799","target":"concept_t15","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2412.03572","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2501.16411","target":"concept_t12","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2406.02347","target":"concept_t32","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2106.08261","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2512.04513","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2109.06780","target":"concept_t22","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2407.18003","target":"concept_t18","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2503.20020","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2406.06973","target":"concept_t23","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2106.02039","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2105.05233","target":"concept_t32","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2412.00887","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2603.07670","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2605.16689","target":"concept_t39","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2506.09042","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2202.01682","target":"concept_t28","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2505.11329","target":"concept_t24","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"oasis-model.github.io:525053739707","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2006.15134","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2502.18080","target":"concept_t27","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2405.06211","target":"concept_t31","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1805.09176","target":"concept_t5","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2509.04633","target":"concept_t5","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2404.16767","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"deepmind.google:f020e8dd5471","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2407.06508","target":"concept_t13","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2410.02664","target":"concept_t22","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2403.02622","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2502.20694","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1711.00832","target":"concept_t22","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2010.02193","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2504.19516","target":"concept_t24","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2312.14125","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2311.13549","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2509.20328","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"philarchive.org:69d8d2ad404e","target":"concept_t2","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"nature.com:413d00acfd43","target":"concept_t28","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.04978","target":"concept_t7","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2605.22882","target":"concept_t15","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2308.07870","target":"concept_t20","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"philarchive.org:04bcc5cbd425","target":"concept_t2","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"x.com:351692e95fe5","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1012.5151","target":"concept_t36","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2402.09664","target":"concept_t27","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2504.05352","target":"concept_t37","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1906.10184","target":"concept_t25","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2210.02747","target":"concept_t3","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2403.04121","target":"concept_t27","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2105.07308","target":"concept_t5","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2603.18532","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2402.01528","target":"concept_t9","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2501.18837","target":"concept_t8","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2109.01134","target":"concept_t23","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2506.21539","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.19818","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2411.13112","target":"concept_t12","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1706.03762","target":"concept_t16","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2307.15818","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2212.13472","target":"concept_t4","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"nature.com:e173b62796c2","target":"concept_t28","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2505.19867","target":"concept_t14","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"openreview.net:3f48a33ff9bf","target":"concept_t14","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2503.16419","target":"concept_t27","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2602.08024","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2410.10377","target":"concept_t39","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2405.08295","target":"concept_t11","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2306.09265","target":"concept_t30","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2002.04501","target":"concept_t25","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2309.13638","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1903.11027","target":"concept_t12","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2402.12275","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2410.07484","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2001.10371","target":"concept_t4","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"1x.tech:d48ab282348b","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2509.06266","target":"concept_t12","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2410.11767","target":"concept_t16","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2602.12520","target":"concept_t22","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2606.15874","target":"concept_t35","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2511.22904","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"philpapers.org:e7fc6fd114b8","target":"concept_t2","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2402.08268","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2601.17067","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2402.06590","target":"concept_t5","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2211.10831","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2403.14403","target":"concept_t31","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2411.04118","target":"concept_t13","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2312.15796","target":"concept_t34","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2311.17918","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1911.10601","target":"concept_t14","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2012.04293","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2406.07507","target":"concept_t3","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2603.20927","target":"concept_t14","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2508.08762","target":"concept_t20","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"nature.com:c3aa5b0d9273","target":"concept_t28","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2511.15722","target":"concept_t12","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2402.02716","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2311.03658","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1910.01442","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2606.09032","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2406.06385","target":"concept_t37","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2211.10438","target":"concept_t37","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2305.16291","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2404.00610","target":"concept_t31","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2105.04906","target":"concept_t23","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"fi.ee.tsinghua.edu.cn:1c1fbee23e33","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2502.11946","target":"concept_t11","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"technologyreview.com:6e6b17c22c9b","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2309.09777","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2506.22567","target":"concept_t13","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2411.15594","target":"concept_t33","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1805.07526","target":"concept_t20","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2402.12451","target":"concept_t30","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2206.08853","target":"concept_t22","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2306.15668","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2403.08763","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2406.07472","target":"concept_t15","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2410.00161","target":"concept_t18","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:0910.3570","target":"concept_t36","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2210.13382","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2308.01399","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"philpapers.org:f2d0d9f7c8fe","target":"concept_t2","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2509.20021","target":"concept_t7","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1711.00937","target":"concept_t20","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2410.08893","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:cs/0405001","target":"concept_t10","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"pi.website:ca5c1a691358","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1909.10863","target":"concept_t14","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2304.07193","target":"concept_t23","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.17111","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2603.21354","target":"concept_t39","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2511.02097","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.25760","target":"concept_t12","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2505.14146","target":"concept_t31","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2303.07109","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2506.01622","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2303.03378","target":"concept_t0","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2406.15836","target":"concept_t22","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2305.14909","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2406.06485","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2504.15785","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2309.07870","target":"concept_t35","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2201.11903","target":"concept_t27","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2603.03251","target":"concept_t9","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2507.05169","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2307.06281","target":"concept_t30","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2412.14161","target":"concept_t35","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2307.12698","target":"concept_t23","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1811.04551","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2103.00020","target":"concept_t23","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2506.18701","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2503.15168","target":"concept_t7","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.19788","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"nature.com:ef956cce5b54","target":"concept_t28","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2306.10754","target":"concept_t4","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2102.11107","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2411.13157","target":"concept_t9","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2001.11841","target":"concept_t14","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2412.17254","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2606.10749","target":"concept_t8","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2407.02485","target":"concept_t31","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2503.02854","target":"concept_t16","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2406.03520","target":"concept_t29","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2604.22748","target":"concept_t38","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2101.03961","target":"concept_t16","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2310.09615","target":"concept_t26","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2402.10644","target":"concept_t16","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2306.02572","target":"concept_t14","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2406.01506","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2312.10997","target":"concept_t31","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1911.03537","target":"concept_t36","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2301.08861","target":"concept_t4","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2111.06377","target":"concept_t23","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2311.17593","target":"concept_t17","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2601.01885","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2005.11401","target":"concept_t31","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2305.17888","target":"concept_t37","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2510.08531","target":"concept_t12","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:1803.07616","target":"concept_t6","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2305.16653","target":"concept_t19","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2507.02768","target":"concept_t11","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2605.12957","target":"concept_t15","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2305.11169","target":"concept_t1","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2409.10644","target":"concept_t9","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2311.17042","target":"concept_t32","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"arxiv:2508.16745","target":"concept_t27","relation":"in-theme","confidence":"INFERRED","confidence_score":0.6},{"source":"concept_t0","target":"concept_t17","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t0","target":"concept_t12","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t0","target":"concept_t26","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.69},{"source":"concept_t1","target":"concept_t27","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t1","target":"concept_t33","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t1","target":"concept_t16","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t2","target":"concept_t5","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.63},{"source":"concept_t2","target":"concept_t8","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.62},{"source":"concept_t2","target":"concept_t6","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.61},{"source":"concept_t3","target":"concept_t32","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t3","target":"concept_t14","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.67},{"source":"concept_t3","target":"concept_t26","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.66},{"source":"concept_t4","target":"concept_t14","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.65},{"source":"concept_t4","target":"concept_t25","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.64},{"source":"concept_t4","target":"concept_t26","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.63},{"source":"concept_t5","target":"concept_t28","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t5","target":"concept_t14","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.69},{"source":"concept_t5","target":"concept_t25","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t6","target":"concept_t27","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t6","target":"concept_t14","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.67},{"source":"concept_t6","target":"concept_t17","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.67},{"source":"concept_t7","target":"concept_t0","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t7","target":"concept_t17","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t7","target":"concept_t38","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.67},{"source":"concept_t8","target":"concept_t19","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t8","target":"concept_t9","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t8","target":"concept_t27","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.67},{"source":"concept_t9","target":"concept_t18","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.72},{"source":"concept_t9","target":"concept_t27","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t9","target":"concept_t37","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t10","target":"concept_t1","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.62},{"source":"concept_t10","target":"concept_t11","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.61},{"source":"concept_t10","target":"concept_t33","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.61},{"source":"concept_t11","target":"concept_t1","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t11","target":"concept_t9","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.67},{"source":"concept_t11","target":"concept_t30","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.66},{"source":"concept_t12","target":"concept_t30","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t12","target":"concept_t17","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t12","target":"concept_t0","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t13","target":"concept_t30","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.67},{"source":"concept_t13","target":"concept_t23","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.66},{"source":"concept_t13","target":"concept_t28","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.66},{"source":"concept_t14","target":"concept_t25","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t14","target":"concept_t26","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t14","target":"concept_t17","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t15","target":"concept_t17","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t15","target":"concept_t29","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.69},{"source":"concept_t15","target":"concept_t12","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.69},{"source":"concept_t16","target":"concept_t9","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t16","target":"concept_t1","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t16","target":"concept_t37","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t17","target":"concept_t0","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t17","target":"concept_t12","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t17","target":"concept_t38","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t18","target":"concept_t24","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.72},{"source":"concept_t18","target":"concept_t37","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.72},{"source":"concept_t18","target":"concept_t9","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.72},{"source":"concept_t19","target":"concept_t27","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t19","target":"concept_t22","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t19","target":"concept_t35","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t20","target":"concept_t28","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t20","target":"concept_t14","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t20","target":"concept_t26","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t21","target":"concept_t34","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.64},{"source":"concept_t21","target":"concept_t17","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.64},{"source":"concept_t21","target":"concept_t38","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.63},{"source":"concept_t22","target":"concept_t26","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t22","target":"concept_t19","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t22","target":"concept_t14","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t23","target":"concept_t30","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t23","target":"concept_t32","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t23","target":"concept_t29","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t24","target":"concept_t18","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.72},{"source":"concept_t24","target":"concept_t37","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t24","target":"concept_t9","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t25","target":"concept_t14","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t25","target":"concept_t5","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t25","target":"concept_t20","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.66},{"source":"concept_t26","target":"concept_t14","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t26","target":"concept_t22","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t26","target":"concept_t17","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t27","target":"concept_t9","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t27","target":"concept_t19","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t27","target":"concept_t1","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t28","target":"concept_t5","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t28","target":"concept_t20","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t28","target":"concept_t14","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.67},{"source":"concept_t29","target":"concept_t15","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.69},{"source":"concept_t29","target":"concept_t17","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.69},{"source":"concept_t29","target":"concept_t32","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.69},{"source":"concept_t30","target":"concept_t12","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t30","target":"concept_t23","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t30","target":"concept_t1","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t31","target":"concept_t27","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t31","target":"concept_t1","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t31","target":"concept_t18","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.69},{"source":"concept_t32","target":"concept_t29","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.69},{"source":"concept_t32","target":"concept_t17","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t32","target":"concept_t23","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t33","target":"concept_t27","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t33","target":"concept_t1","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t33","target":"concept_t31","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.69},{"source":"concept_t34","target":"concept_t17","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.66},{"source":"concept_t34","target":"concept_t32","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.65},{"source":"concept_t34","target":"concept_t9","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.65},{"source":"concept_t35","target":"concept_t19","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.7},{"source":"concept_t35","target":"concept_t33","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.67},{"source":"concept_t35","target":"concept_t22","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.67},{"source":"concept_t36","target":"concept_t25","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.64},{"source":"concept_t36","target":"concept_t20","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.63},{"source":"concept_t36","target":"concept_t32","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.63},{"source":"concept_t37","target":"concept_t18","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.72},{"source":"concept_t37","target":"concept_t24","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t37","target":"concept_t9","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t38","target":"concept_t17","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.71},{"source":"concept_t38","target":"concept_t0","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t38","target":"concept_t15","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.68},{"source":"concept_t39","target":"concept_t24","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.65},{"source":"concept_t39","target":"concept_t16","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.64},{"source":"concept_t39","target":"concept_t9","relation":"adjacent-theme","confidence":"INFERRED","confidence_score":0.64},{"source":"arxiv:1708.00133","target":"arxiv:2101.07393","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:1708.00133","target":"doi.org:0a734c9c22c9","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1708.00133","target":"arxiv:2410.02664","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1708.00133","target":"arxiv:1910.08210","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2603.00479","target":"arxiv:2512.00597","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2603.00479","target":"arxiv:2210.04133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2603.00479","target":"arxiv:2501.02189","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2603.00479","target":"arxiv:2404.07214","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2305.10601","target":"arxiv:2305.14992","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2305.10601","target":"arxiv:2201.11903","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2305.10601","target":"arxiv:2205.11916","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2305.10601","target":"arxiv:2502.18080","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2502.12130","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2502.12130","target":"arxiv:2509.02547","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2502.12130","target":"arxiv:2402.02716","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2502.12130","target":"arxiv:2412.14161","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2510.26802","target":"arxiv:2509.20328","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2510.26802","target":"arxiv:2601.17067","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2510.26802","target":"arxiv:2501.09038","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2510.26802","target":"arxiv:2603.28489","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2510.21219","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2510.21219","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2510.21219","target":"arxiv:2411.14499","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2510.21219","target":"technologyreview.com:6e6b17c22c9b","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2312.03815","target":"arxiv:2403.16971","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2312.03815","target":"arxiv:2309.07870","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2312.03815","target":"arxiv:2308.03688","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2312.03815","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2402.01695","target":"arxiv:2511.22904","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2402.01695","target":"arxiv:2305.10626","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2402.01695","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2402.01695","target":"arxiv:2308.01399","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2311.05348","target":"arxiv:2402.12451","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2311.05348","target":"arxiv:2310.03744","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2311.05348","target":"arxiv:2411.06284","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2311.05348","target":"arxiv:2501.02189","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2401.12945","target":"arxiv:2204.03458","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2401.12945","target":"arxiv:2311.15127","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2401.12945","target":"arxiv:2312.14125","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2401.12945","target":"arxiv:2401.09985","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2504.00983","target":"arxiv:2502.20694","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2504.00983","target":"arxiv:2410.18072","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2504.00983","target":"arxiv:2401.09985","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2504.00983","target":"openreview.net:c15700430713","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:1910.10683","target":"arxiv:2005.14165","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1910.10683","target":"arxiv:2403.08540","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:1910.10683","target":"arxiv:2403.08763","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:1910.10683","target":"arxiv:2103.00020","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2512.04797","target":"deepmind.google:f020e8dd5471","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2512.04797","target":"deepmind.google:39914de6d3c9","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2512.04797","target":"arxiv:2503.20020","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2512.04797","target":"arxiv:2507.00917","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"psychclassics.yorku.ca:d7323f879ff8","target":"arxiv:2202.01682","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"psychclassics.yorku.ca:d7323f879ff8","target":"nature.com:c3aa5b0d9273","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"psychclassics.yorku.ca:d7323f879ff8","target":"pubmed.ncbi.nlm.nih.gov:fd66789ead66","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"psychclassics.yorku.ca:d7323f879ff8","target":"frontiersin.org:d1ec3c5c2c89","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2311.07222","target":"nature.com:386c3df42560","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2311.07222","target":"arxiv:2406.14798","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2311.07222","target":"arxiv:2212.12794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2311.07222","target":"arxiv:2312.15796","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2605.12090","target":"arxiv:2510.16732","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2605.12090","target":"arxiv:2602.22010","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2605.12090","target":"arxiv:2506.21539","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2605.12090","target":"fi.ee.tsinghua.edu.cn:1c1fbee23e33","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2506.18897","target":"arxiv:2602.10717","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2506.18897","target":"arxiv:2601.21998","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2506.18897","target":"arxiv:2511.08585","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2506.18897","target":"arxiv:2605.06192","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2511.00062","target":"huggingface.co:e9042540da12","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.94},{"source":"arxiv:2511.00062","target":"arxiv:2501.03575","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2511.00062","target":"blogs.nvidia.com:9f8a6e618fbd","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2511.00062","target":"arxiv:2410.18072","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2601.15533","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2601.15533","target":"arxiv:2511.08585","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2601.15533","target":"arxiv:2507.05169","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2601.15533","target":"arxiv:2601.17067","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2510.04999","target":"arxiv:2204.03458","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2510.04999","target":"arxiv:2305.18264","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2510.04999","target":"arxiv:2603.28489","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2510.04999","target":"arxiv:2311.15127","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2304.12995","target":"arxiv:2502.11946","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2304.12995","target":"arxiv:2506.08967","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2304.12995","target":"arxiv:2102.01192","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2304.12995","target":"arxiv:2405.08295","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2602.22010","target":"arxiv:2605.12090","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2602.22010","target":"arxiv:2602.10717","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2602.22010","target":"arxiv:2510.19818","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2602.22010","target":"arxiv:2506.21539","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2403.15498","target":"arxiv:2210.13382","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2403.15498","target":"arxiv:2309.00941","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2403.15498","target":"arxiv:2605.23972","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2403.15498","target":"arxiv:2406.06485","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2510.05865","target":"arxiv:2411.07690","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2510.05865","target":"arxiv:2510.16732","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2510.05865","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2510.05865","target":"arxiv:2403.02622","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"proceedings.mlr.press:6beb908867f9","target":"arxiv:2305.11169","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"proceedings.mlr.press:6beb908867f9","target":"arxiv:2402.09664","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"proceedings.mlr.press:6beb908867f9","target":"arxiv:2509.14252","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"proceedings.mlr.press:6beb908867f9","target":"arxiv:2305.17926","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2310.02207","target":"arxiv:2310.04276","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2310.02207","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2310.02207","target":"arxiv:2106.00737","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2310.02207","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2603.27918","target":"arxiv:2402.10196","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2603.27918","target":"arxiv:2407.15549","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2603.27918","target":"arxiv:2605.17288","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2603.27918","target":"arxiv:2411.06284","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2602.15400","target":"arxiv:2508.13073","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2602.15400","target":"arxiv:2411.13112","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2602.15400","target":"arxiv:2510.25760","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2602.15400","target":"arxiv:2510.17111","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2403.06845","target":"arxiv:2309.09777","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2403.06845","target":"arxiv:2411.10171","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2403.06845","target":"arxiv:2502.10498","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2403.06845","target":"arxiv:2506.09042","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2507.00917","target":"arxiv:2510.16732","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2507.00917","target":"arxiv:2407.06886","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2507.00917","target":"arxiv:2509.20021","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2507.00917","target":"arxiv:2506.22355","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2402.15391","target":"deepmind.google:f020e8dd5471","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2402.15391","target":"deepmind.google:39914de6d3c9","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2402.15391","target":"arxiv:2511.08585","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2402.15391","target":"arxiv:2406.09455","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2102.01192","target":"arxiv:2309.05767","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2102.01192","target":"arxiv:2304.12995","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2102.01192","target":"arxiv:2405.08295","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2102.01192","target":"arxiv:2506.08967","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:1705.09156","target":"arxiv:2108.13343","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:1705.09156","target":"nature.com:7fc0dd8e4038","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:1705.09156","target":"arxiv:2201.06387","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:1705.09156","target":"arxiv:2105.11203","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:1706.03741","target":"doi.org:0a734c9c22c9","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1706.03741","target":"arxiv:2007.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1706.03741","target":"arxiv:2208.01769","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1706.03741","target":"arxiv:2106.02039","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2204.03458","target":"arxiv:2311.15127","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2204.03458","target":"arxiv:2105.05233","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2204.03458","target":"arxiv:2112.10752","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2204.03458","target":"arxiv:2510.04999","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2507.10535","target":"arxiv:2411.15594","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2507.10535","target":"arxiv:2402.09664","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2507.10535","target":"arxiv:2306.05685","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2507.10535","target":"arxiv:2305.17926","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2005.14165","target":"arxiv:2205.11916","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2005.14165","target":"arxiv:2403.08540","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2005.14165","target":"arxiv:2401.02954","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2005.14165","target":"arxiv:1910.10683","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2512.00597","target":"arxiv:2210.04133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2512.00597","target":"arxiv:2603.00479","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2512.00597","target":"arxiv:2411.04118","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2512.00597","target":"arxiv:2506.22567","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2210.03629","target":"arxiv:2305.14992","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2210.03629","target":"arxiv:2303.11366","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2210.03629","target":"arxiv:2201.11903","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2210.03629","target":"arxiv:2205.11916","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2202.09467","target":"arxiv:2308.07870","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2202.09467","target":"arxiv:2103.03725","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2202.09467","target":"arxiv:2006.04182","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2202.09467","target":"arxiv:2508.08762","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2410.02548","target":"arxiv:2210.02747","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2410.02548","target":"arxiv:2406.07507","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2410.02548","target":"arxiv:2304.14772","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2410.02548","target":"arxiv:2504.10612","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2307.16125","target":"arxiv:2307.06281","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2307.16125","target":"arxiv:2411.06284","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2307.16125","target":"arxiv:2402.12451","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2307.16125","target":"arxiv:2502.14739","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2510.20416","target":"arxiv:2311.07222","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2510.20416","target":"arxiv:2212.12794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2510.20416","target":"arxiv:2306.14066","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2510.20416","target":"nature.com:386c3df42560","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2602.10717","target":"arxiv:2601.21998","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2602.10717","target":"arxiv:2602.22010","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2602.10717","target":"arxiv:2506.18897","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2602.10717","target":"arxiv:2511.02097","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2407.08867","target":"arxiv:2310.08803","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2407.08867","target":"arxiv:2408.03943","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2407.08867","target":"arxiv:2510.04978","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2407.08867","target":"arxiv:2503.15168","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2309.02591","target":"arxiv:2402.12451","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2309.02591","target":"arxiv:2310.03744","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2309.02591","target":"arxiv:2110.04544","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2309.02591","target":"arxiv:2109.01134","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2503.06800","target":"arxiv:2406.03520","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"arxiv:2503.06800","target":"arxiv:2501.09038","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2503.06800","target":"arxiv:2411.02385","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2503.06800","target":"arxiv:2502.20694","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2510.03847","target":"arxiv:2308.03688","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2510.03847","target":"arxiv:2603.20224","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2510.03847","target":"arxiv:2503.16416","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2510.03847","target":"arxiv:2401.02954","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"openreview.net:c15700430713","target":"arxiv:2406.07472","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"openreview.net:c15700430713","target":"arxiv:2311.01017","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"openreview.net:c15700430713","target":"arxiv:2606.18250","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"openreview.net:c15700430713","target":"arxiv:2411.10171","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2207.12316","target":"arxiv:2103.03725","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2207.12316","target":"arxiv:2202.09467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2207.12316","target":"arxiv:2508.08762","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2207.12316","target":"arxiv:2006.04182","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2303.01469","target":"arxiv:2105.05233","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2303.01469","target":"arxiv:2311.17042","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2303.01469","target":"arxiv:2204.03458","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2303.01469","target":"arxiv:2406.07507","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"transformer-circuits.pub:e2306116790e","target":"arxiv:2309.08600","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"transformer-circuits.pub:e2306116790e","target":"arxiv:2406.11944","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"transformer-circuits.pub:e2306116790e","target":"arxiv:2202.08906","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"transformer-circuits.pub:e2306116790e","target":"arxiv:2410.11767","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:1908.05656","target":"arxiv:2106.08261","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:1908.05656","target":"arxiv:2501.16411","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:1908.05656","target":"arxiv:2306.15668","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:1908.05656","target":"arxiv:1803.07616","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2410.18072","target":"arxiv:2502.20694","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2410.18072","target":"arxiv:2511.08585","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2410.18072","target":"arxiv:2510.16732","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2410.18072","target":"arxiv:2511.00062","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"cambridge.org:57359c90d672","target":"arxiv:2305.10601","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"cambridge.org:57359c90d672","target":"arxiv:2210.13382","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"cambridge.org:57359c90d672","target":"arxiv:2310.08803","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"cambridge.org:57359c90d672","target":"arxiv:2604.16592","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.76},{"source":"nature.com:7719482cd2c4","target":"arxiv:1712.04602","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"nature.com:7719482cd2c4","target":"frontiersin.org:d1ec3c5c2c89","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"nature.com:7719482cd2c4","target":"arxiv:1805.09042","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"nature.com:7719482cd2c4","target":"pubmed.ncbi.nlm.nih.gov:fd66789ead66","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"transformer-circuits.pub:c4d9a1f0e2b7","target":"arxiv:2506.21876","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"transformer-circuits.pub:c4d9a1f0e2b7","target":"arxiv:2307.15054","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"transformer-circuits.pub:c4d9a1f0e2b7","target":"arxiv:2310.04276","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"transformer-circuits.pub:c4d9a1f0e2b7","target":"arxiv:2503.15168","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2308.03688","target":"arxiv:2401.13178","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2308.03688","target":"arxiv:2503.16416","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2308.03688","target":"arxiv:2412.14161","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2308.03688","target":"neurips.cc:9c7e4f09bb35","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2309.17080","target":"arxiv:2503.20523","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"arxiv:2309.17080","target":"arxiv:2403.02622","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2309.17080","target":"arxiv:2405.17398","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2309.17080","target":"arxiv:2405.03520","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"nature.com:7fc0dd8e4038","target":"arxiv:1705.09156","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"nature.com:7fc0dd8e4038","target":"arxiv:2201.06387","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"nature.com:7fc0dd8e4038","target":"arxiv:2112.15242","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"nature.com:7fc0dd8e4038","target":"arxiv:2108.13343","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2606.24467","target":"arxiv:2410.00161","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2606.24467","target":"arxiv:2412.02252","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2606.24467","target":"arxiv:2508.06297","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2606.24467","target":"arxiv:2509.00579","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2604.06882","target":"arxiv:2605.16689","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2604.06882","target":"arxiv:2605.16547","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2604.06882","target":"arxiv:2505.01712","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2604.06882","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:1803.10122","target":"people.idsia.ch:73d35fac50f9","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:1803.10122","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1803.10122","target":"arxiv:2511.22904","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1803.10122","target":"arxiv:1912.01603","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2502.11831","target":"arxiv:2506.09849","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2502.11831","target":"arxiv:2501.09038","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2502.11831","target":"arxiv:2404.08471","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2502.11831","target":"arxiv:1803.07616","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2406.14798","target":"arxiv:2306.14066","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2406.14798","target":"arxiv:2311.07222","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2406.14798","target":"nature.com:386c3df42560","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2406.14798","target":"arxiv:2312.15796","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1312.6114","target":"arxiv:1711.00937","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:1312.6114","target":"arxiv:1709.02341","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:1312.6114","target":"arxiv:2006.11239","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:1312.6114","target":"arxiv:1907.05600","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2312.03788","target":"arxiv:2211.10438","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.94},{"source":"arxiv:2312.03788","target":"arxiv:2308.09723","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2312.03788","target":"arxiv:2305.17888","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2312.03788","target":"arxiv:2504.05352","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2407.00118","target":"arxiv:2402.12451","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2407.00118","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2407.00118","target":"arxiv:2511.02097","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2407.00118","target":"arxiv:2411.14499","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1709.06560","target":"arxiv:2107.08241","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:1709.06560","target":"arxiv:2007.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1709.06560","target":"arxiv:2006.15134","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1709.06560","target":"arxiv:1706.03741","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2204.00227","target":"arxiv:2006.11239","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2204.00227","target":"arxiv:2105.05233","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2204.00227","target":"arxiv:2303.01469","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2204.00227","target":"arxiv:2311.17042","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2305.10626","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2305.10626","target":"arxiv:2402.01695","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2305.10626","target":"arxiv:2511.22904","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2305.10626","target":"arxiv:2308.01399","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2506.09985","target":"arxiv:2404.08471","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2506.09985","target":"arxiv:2307.12698","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2506.09985","target":"arxiv:2602.10717","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2506.09985","target":"openreview.net:3f48a33ff9bf","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2110.04544","target":"arxiv:2109.01134","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2110.04544","target":"arxiv:2406.06973","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2110.04544","target":"arxiv:2305.07017","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2110.04544","target":"arxiv:2412.07077","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2401.15884","target":"arxiv:2312.10997","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2401.15884","target":"arxiv:2404.00610","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2401.15884","target":"arxiv:2405.06211","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2401.15884","target":"arxiv:2005.11401","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2501.09038","target":"arxiv:2502.11831","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2501.09038","target":"arxiv:2406.03520","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2501.09038","target":"arxiv:2411.02385","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2501.09038","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2502.05855","target":"arxiv:2510.17111","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2502.05855","target":"arxiv:2507.12440","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2502.05855","target":"arxiv:2508.13073","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2502.05855","target":"arxiv:2603.18532","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2310.08560","target":"arxiv:2603.07670","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2310.08560","target":"arxiv:2601.01885","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2310.08560","target":"arxiv:2407.12036","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2310.08560","target":"arxiv:2403.04121","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2305.15074","target":"arxiv:2307.13692","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2305.15074","target":"arxiv:2502.14739","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2305.15074","target":"arxiv:2506.21521","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2305.15074","target":"arxiv:2411.15594","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2602.24044","target":"arxiv:2504.19516","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2602.24044","target":"arxiv:2504.09590","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2602.24044","target":"arxiv:2603.20224","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2602.24044","target":"arxiv:2503.20552","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2503.02143","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2503.02143","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2503.02143","target":"arxiv:2510.04978","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2503.02143","target":"arxiv:2503.15168","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2103.03725","target":"arxiv:2207.12316","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2103.03725","target":"arxiv:2202.09467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2103.03725","target":"arxiv:2006.04182","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2103.03725","target":"arxiv:2508.08762","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2510.12966","target":"arxiv:2409.10644","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2510.12966","target":"arxiv:2402.01528","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2510.12966","target":"arxiv:2510.13161","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2510.12966","target":"arxiv:2511.20340","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2604.16592","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2604.16592","target":"arxiv:2503.15168","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2604.16592","target":"arxiv:2509.20021","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2604.16592","target":"arxiv:2408.03943","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2605.02900","target":"arxiv:2510.05865","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2605.02900","target":"arxiv:2407.06886","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2605.02900","target":"arxiv:2507.00917","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2605.02900","target":"arxiv:2509.20021","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2505.21996","target":"arxiv:2603.28489","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2505.21996","target":"arxiv:2601.17067","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2505.21996","target":"arxiv:2605.06192","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2505.21996","target":"huggingface.co:9f24e732f5fe","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2507.05201","target":"arxiv:2411.04118","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2507.05201","target":"arxiv:2303.00915","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2507.05201","target":"arxiv:2210.04133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2507.05201","target":"arxiv:2506.22567","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2202.08906","target":"arxiv:2101.03961","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2202.08906","target":"transformer-circuits.pub:e2306116790e","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2202.08906","target":"arxiv:2406.11944","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2202.08906","target":"arxiv:2403.08540","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2308.12284","target":"arxiv:2403.08763","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2308.12284","target":"arxiv:2502.21321","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2308.12284","target":"arxiv:2509.14252","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2308.12284","target":"arxiv:2401.02954","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2208.01769","target":"arxiv:2509.02547","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2208.01769","target":"arxiv:1711.00832","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2208.01769","target":"arxiv:1812.00922","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2208.01769","target":"arxiv:1706.02275","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2512.02419","target":"arxiv:2308.07870","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2512.02419","target":"arxiv:2402.06590","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2512.02419","target":"nature.com:a4d8ac562091","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2512.02419","target":"arxiv:2503.15168","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2603.28489","target":"arxiv:2601.17067","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2603.28489","target":"arxiv:2511.08585","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2603.28489","target":"arxiv:2505.21996","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2603.28489","target":"arxiv:2506.18897","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2511.12239","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2511.12239","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2511.12239","target":"arxiv:2503.15168","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2511.12239","target":"arxiv:2510.21219","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2502.03373","target":"arxiv:2502.18080","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2502.03373","target":"arxiv:2503.16419","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2502.03373","target":"arxiv:2502.21321","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2502.03373","target":"arxiv:2205.11916","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:1011.0063","target":"arxiv:1012.5151","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"arxiv:1011.0063","target":"arxiv:1911.03537","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1011.0063","target":"arxiv:0910.3570","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1011.0063","target":"arxiv:2208.06726","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2605.17288","target":"arxiv:2407.15549","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2605.17288","target":"arxiv:2603.27918","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2605.17288","target":"arxiv:2402.10196","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2605.17288","target":"arxiv:2309.13638","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:1712.04602","target":"arxiv:1805.09042","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:1712.04602","target":"nature.com:7719482cd2c4","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:1712.04602","target":"frontiersin.org:d1ec3c5c2c89","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1712.04602","target":"nature.com:ad56347f1921","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2504.09590","target":"arxiv:2504.19516","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2504.09590","target":"arxiv:2412.18106","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2504.09590","target":"arxiv:2602.24044","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2504.09590","target":"arxiv:2603.20397","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2605.29930","target":"arxiv:2510.21219","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2605.29930","target":"arxiv:2503.15168","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2605.29930","target":"arxiv:2510.04978","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2605.29930","target":"arxiv:2511.12239","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2412.03568","target":"arxiv:2506.18701","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2412.03568","target":"arxiv:2310.06114","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2412.03568","target":"arxiv:2603.28489","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2412.03568","target":"arxiv:2511.08585","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2206.04114","target":"arxiv:1811.04551","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2206.04114","target":"openreview.net:3f48a33ff9bf","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2206.04114","target":"arxiv:1706.03741","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2206.04114","target":"arxiv:2107.08241","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2411.17470","target":"arxiv:2212.09748","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2411.17470","target":"arxiv:2204.03458","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2411.17470","target":"arxiv:2311.15127","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2411.17470","target":"arxiv:2411.02385","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2411.07690","target":"arxiv:2510.05865","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2411.07690","target":"arxiv:2403.02622","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2411.07690","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2411.07690","target":"arxiv:2411.14499","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1907.05600","target":"arxiv:2006.11239","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:1907.05600","target":"arxiv:2105.05233","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1907.05600","target":"arxiv:2303.01469","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1907.05600","target":"arxiv:2204.00227","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2510.01539","target":"arxiv:2502.21321","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2510.01539","target":"arxiv:2508.16745","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2510.01539","target":"arxiv:2502.03373","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2510.01539","target":"arxiv:2503.16419","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2601.21998","target":"arxiv:2602.10717","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2601.21998","target":"arxiv:2506.18897","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2601.21998","target":"arxiv:2605.06192","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2601.21998","target":"arxiv:2606.06556","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2305.17198","target":"arxiv:1711.00832","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2305.17198","target":"arxiv:2602.12520","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2305.17198","target":"arxiv:2208.01769","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2305.17198","target":"arxiv:1812.00922","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2404.12377","target":"arxiv:2602.10717","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2404.12377","target":"arxiv:2511.08585","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2404.12377","target":"arxiv:2606.06556","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2404.12377","target":"arxiv:2601.21998","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2210.07729","target":"arxiv:2411.10171","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2210.07729","target":"arxiv:2403.02622","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2210.07729","target":"arxiv:2502.10498","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2210.07729","target":"arxiv:2311.01017","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2209.00588","target":"arxiv:2310.09615","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2209.00588","target":"arxiv:2303.07109","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2209.00588","target":"arxiv:2406.15836","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2209.00588","target":"arxiv:2010.02193","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2305.18264","target":"arxiv:2204.03458","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2305.18264","target":"arxiv:2412.17254","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2305.18264","target":"arxiv:2311.15127","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2305.18264","target":"arxiv:2510.04999","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2605.23972","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2605.23972","target":"arxiv:2406.06485","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2605.23972","target":"arxiv:2511.22904","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2605.23972","target":"arxiv:2305.14992","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2407.15549","target":"arxiv:2605.17288","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2407.15549","target":"arxiv:2502.21321","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2407.15549","target":"arxiv:2603.27918","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2407.15549","target":"arxiv:2501.18837","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"frontiersin.org:d1ec3c5c2c89","target":"arxiv:2202.01682","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"frontiersin.org:d1ec3c5c2c89","target":"arxiv:1805.09042","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"frontiersin.org:d1ec3c5c2c89","target":"arxiv:1712.04602","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"frontiersin.org:d1ec3c5c2c89","target":"nature.com:7719482cd2c4","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2506.09171","target":"arxiv:2411.06559","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2506.09171","target":"arxiv:2402.02716","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2506.09171","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2506.09171","target":"arxiv:2601.01885","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1912.01603","target":"arxiv:2010.02193","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1912.01603","target":"arxiv:2206.14176","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1912.01603","target":"arxiv:1803.10122","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1912.01603","target":"nature.com:f005fae15467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1805.09042","target":"arxiv:1712.04602","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:1805.09042","target":"arxiv:2202.01682","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:1805.09042","target":"nature.com:ef956cce5b54","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1805.09042","target":"frontiersin.org:d1ec3c5c2c89","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2310.01361","target":"arxiv:2603.18532","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2310.01361","target":"arxiv:2310.06114","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2310.01361","target":"arxiv:2606.06556","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2310.01361","target":"arxiv:2404.12377","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2401.13178","target":"arxiv:2308.03688","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2401.13178","target":"arxiv:2503.16416","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2401.13178","target":"arxiv:2606.09032","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2401.13178","target":"arxiv:2412.14161","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2112.07103","target":"arxiv:2103.04723","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2112.07103","target":"arxiv:2301.08861","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2112.07103","target":"arxiv:2306.10754","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2112.07103","target":"arxiv:2212.13472","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2406.00877","target":"arxiv:2111.09259","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2406.00877","target":"arxiv:2403.15498","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2406.00877","target":"arxiv:2309.00941","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2406.00877","target":"nature.com:e56ca063b728","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"nature.com:e56ca063b728","target":"arxiv:1911.08265","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"nature.com:e56ca063b728","target":"doi.org:0a734c9c22c9","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"nature.com:e56ca063b728","target":"arxiv:2111.09259","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"nature.com:e56ca063b728","target":"arxiv:2406.00877","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2310.08803","target":"arxiv:1807.11819","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2310.08803","target":"arxiv:2503.15168","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2310.08803","target":"arxiv:2510.04978","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2310.08803","target":"arxiv:2512.02419","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2505.02074","target":"arxiv:2102.11107","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2505.02074","target":"arxiv:2506.23068","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2505.02074","target":"arxiv:2601.17067","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2505.02074","target":"arxiv:2509.12387","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2401.07851","target":"arxiv:2411.13157","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"arxiv:2401.07851","target":"arxiv:2402.01528","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2401.07851","target":"arxiv:2511.20340","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2401.07851","target":"arxiv:2603.03251","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2309.17382","target":"arxiv:2305.14992","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2309.17382","target":"arxiv:2509.02547","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2309.17382","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2309.17382","target":"arxiv:2605.23972","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2010.11929","target":"arxiv:2212.09748","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2010.11929","target":"arxiv:1706.03762","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2010.11929","target":"arxiv:2405.10480","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2010.11929","target":"arxiv:2209.00588","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2506.23068","target":"arxiv:2505.02074","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2506.23068","target":"arxiv:2509.12387","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2506.23068","target":"arxiv:2102.11107","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2506.23068","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2508.13073","target":"arxiv:2510.17111","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2508.13073","target":"arxiv:2307.15818","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2508.13073","target":"arxiv:2502.05855","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2508.13073","target":"arxiv:2507.12440","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2406.16062","target":"arxiv:2508.08762","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2406.16062","target":"arxiv:2006.04182","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2406.16062","target":"arxiv:2103.03725","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2406.16062","target":"arxiv:2312.09257","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2212.09748","target":"arxiv:2411.17470","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2212.09748","target":"arxiv:2105.05233","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2212.09748","target":"arxiv:2204.03458","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2212.09748","target":"arxiv:2112.10752","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2401.02954","target":"arxiv:2502.21321","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2401.02954","target":"arxiv:2403.08540","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2401.02954","target":"arxiv:2403.08763","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2401.02954","target":"arxiv:2502.14739","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2310.06114","target":"arxiv:2511.08585","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2310.06114","target":"arxiv:2603.18532","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2310.06114","target":"arxiv:2605.16395","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2310.06114","target":"arxiv:2506.18897","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2206.10498","target":"arxiv:2305.14909","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2206.10498","target":"arxiv:2403.04121","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2206.10498","target":"arxiv:2305.14992","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2206.10498","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2305.17926","target":"arxiv:2306.05685","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2305.17926","target":"arxiv:2411.15594","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2305.17926","target":"arxiv:2507.10535","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2305.17926","target":"arxiv:2502.14739","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2606.20545","target":"arxiv:2601.17067","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2606.20545","target":"arxiv:2502.20694","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2606.20545","target":"thesequence.substack.com:0f25bda15c4b","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2606.20545","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2405.10480","target":"arxiv:1706.03762","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2405.10480","target":"arxiv:2402.10644","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2405.10480","target":"arxiv:2101.03961","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2405.10480","target":"arxiv:2511.20340","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2203.03417","target":"arxiv:2202.03771","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2203.03417","target":"arxiv:2306.10754","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2203.03417","target":"arxiv:2208.01769","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2203.03417","target":"arxiv:2406.15836","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2205.06175","target":"arxiv:2509.19958","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2205.06175","target":"arxiv:2206.08853","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2205.06175","target":"arxiv:2304.03442","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2205.06175","target":"arxiv:2503.20523","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2411.06559","target":"arxiv:2506.09171","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2411.06559","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2411.06559","target":"arxiv:2402.02716","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2411.06559","target":"arxiv:2606.09032","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2508.06297","target":"arxiv:2603.20397","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"arxiv:2508.06297","target":"arxiv:2410.00161","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2508.06297","target":"arxiv:2606.24467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2508.06297","target":"arxiv:2407.18003","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2310.08864","target":"arxiv:2212.06817","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2310.08864","target":"arxiv:2307.15818","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2310.08864","target":"pi.website:ca5c1a691358","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2310.08864","target":"arxiv:2606.06556","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2405.12399","target":"arxiv:2408.14837","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2405.12399","target":"arxiv:2010.02193","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2405.12399","target":"arxiv:2407.01392","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2405.12399","target":"arxiv:2105.05233","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2505.01712","target":"arxiv:2605.16547","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2505.01712","target":"arxiv:2403.02622","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2505.01712","target":"arxiv:2604.06882","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2505.01712","target":"arxiv:2411.10171","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2206.14176","target":"arxiv:2010.02193","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2206.14176","target":"arxiv:1912.01603","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2206.14176","target":"arxiv:2602.10717","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2206.14176","target":"nature.com:f005fae15467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2603.20397","target":"arxiv:2508.06297","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"arxiv:2603.20397","target":"arxiv:2605.07234","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2603.20397","target":"arxiv:2407.18003","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2603.20397","target":"arxiv:2509.00579","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2408.14837","target":"arxiv:2405.12399","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2408.14837","target":"arxiv:2412.00887","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2408.14837","target":"arxiv:2407.01392","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2408.14837","target":"arxiv:2105.05233","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2401.09985","target":"arxiv:2502.20694","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2401.09985","target":"arxiv:2506.18897","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2401.09985","target":"arxiv:2410.18072","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2401.09985","target":"arxiv:2506.01799","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2503.14492","target":"huggingface.co:e9042540da12","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2503.14492","target":"arxiv:2511.00062","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2503.14492","target":"github.com:14ee211b0b5d","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2503.14492","target":"blogs.nvidia.com:9f8a6e618fbd","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2007.08794","target":"arxiv:2404.16767","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2007.08794","target":"arxiv:2509.02547","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2007.08794","target":"arxiv:2006.15134","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2007.08794","target":"arxiv:1709.06560","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2210.04133","target":"arxiv:2512.00597","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2210.04133","target":"arxiv:2603.00479","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2210.04133","target":"arxiv:2506.22567","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2210.04133","target":"arxiv:2303.00915","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2506.21876","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2506.21876","target":"arxiv:2404.07214","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2506.21876","target":"arxiv:2510.19818","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2506.21876","target":"arxiv:2411.14499","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2403.08540","target":"arxiv:2403.08763","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2403.08540","target":"arxiv:2401.02954","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2403.08540","target":"arxiv:2502.21321","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2403.08540","target":"arxiv:2005.14165","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2112.15242","target":"arxiv:2108.13343","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2112.15242","target":"arxiv:2201.06387","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2112.15242","target":"arxiv:1906.10184","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2112.15242","target":"arxiv:2105.11203","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2201.06387","target":"arxiv:1906.10184","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.94},{"source":"arxiv:2201.06387","target":"arxiv:2108.13343","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.94},{"source":"arxiv:2201.06387","target":"arxiv:2105.11203","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2201.06387","target":"arxiv:2002.04501","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2309.00941","target":"arxiv:2210.13382","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2309.00941","target":"arxiv:2403.15498","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2309.00941","target":"arxiv:2106.00737","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2309.00941","target":"arxiv:2509.04633","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:1911.11641","target":"arxiv:2510.04978","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:1911.11641","target":"arxiv:2406.03520","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:1911.11641","target":"arxiv:2503.15558","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:1911.11641","target":"arxiv:2012.04293","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"philpapers.org:7b684275d20a","target":"philpapers.org:e7fc6fd114b8","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.94},{"source":"philpapers.org:7b684275d20a","target":"philarchive.org:69d8d2ad404e","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"philpapers.org:7b684275d20a","target":"philarchive.org:04bcc5cbd425","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"philpapers.org:7b684275d20a","target":"philpapers.org:9a9c4dcb81d9","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"transformer-circuits.pub:b904370d3db5","target":"arxiv:2406.11944","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"transformer-circuits.pub:b904370d3db5","target":"arxiv:2407.12036","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"transformer-circuits.pub:b904370d3db5","target":"arxiv:2309.13638","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"transformer-circuits.pub:b904370d3db5","target":"arxiv:2403.04121","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"deepmind.google:39914de6d3c9","target":"deepmind.google:f020e8dd5471","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"deepmind.google:39914de6d3c9","target":"arxiv:2402.15391","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"deepmind.google:39914de6d3c9","target":"arxiv:2406.09455","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"deepmind.google:39914de6d3c9","target":"arxiv:2512.04797","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:1807.11819","target":"arxiv:2312.09257","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:1807.11819","target":"arxiv:2202.01682","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1807.11819","target":"arxiv:2512.02419","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1807.11819","target":"arxiv:2310.08803","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2502.14819","target":"arxiv:2106.02039","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2502.14819","target":"arxiv:1811.04551","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2502.14819","target":"arxiv:2107.08241","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2502.14819","target":"arxiv:2102.08363","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"nature.com:f005fae15467","target":"arxiv:2301.04104","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.96},{"source":"nature.com:f005fae15467","target":"arxiv:2010.02193","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"nature.com:f005fae15467","target":"arxiv:1912.01603","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"nature.com:f005fae15467","target":"arxiv:2206.14176","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2309.05767","target":"arxiv:2102.01192","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2309.05767","target":"arxiv:2405.08295","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2309.05767","target":"arxiv:2507.02768","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2309.05767","target":"arxiv:2506.08967","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2602.03916","target":"arxiv:2511.15722","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2602.03916","target":"arxiv:2509.06266","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2602.03916","target":"arxiv:2510.08531","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2602.03916","target":"arxiv:2510.25760","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2212.12794","target":"nature.com:386c3df42560","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2212.12794","target":"arxiv:2312.15796","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2212.12794","target":"arxiv:2311.07222","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2212.12794","target":"arxiv:2510.20416","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2309.08600","target":"transformer-circuits.pub:e2306116790e","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2309.08600","target":"arxiv:2410.11767","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2309.08600","target":"arxiv:2406.11944","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2309.08600","target":"nature.com:a4d8ac562091","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2502.21321","target":"arxiv:2403.08763","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2502.21321","target":"arxiv:2407.12036","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2502.21321","target":"arxiv:2502.03373","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2502.21321","target":"arxiv:2503.16419","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"thesequence.substack.com:0f25bda15c4b","target":"reply.com:5748177416db","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"thesequence.substack.com:0f25bda15c4b","target":"technologyreview.com:6e6b17c22c9b","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"thesequence.substack.com:0f25bda15c4b","target":"introl.com:c695c416f507","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"thesequence.substack.com:0f25bda15c4b","target":"arxiv:2501.03575","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2312.09257","target":"arxiv:2406.16062","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2312.09257","target":"arxiv:2308.07870","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2312.09257","target":"arxiv:2508.08762","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2312.09257","target":"arxiv:1807.11819","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"reply.com:5748177416db","target":"thesequence.substack.com:0f25bda15c4b","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"reply.com:5748177416db","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"reply.com:5748177416db","target":"technologyreview.com:6e6b17c22c9b","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"reply.com:5748177416db","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"aclanthology.org:8f7acb76ba1d","target":"arxiv:1910.10683","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"aclanthology.org:8f7acb76ba1d","target":"nature.com:a4d8ac562091","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"aclanthology.org:8f7acb76ba1d","target":"arxiv:2005.14165","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"aclanthology.org:8f7acb76ba1d","target":"arxiv:2502.21321","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2112.10752","target":"arxiv:2105.05233","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2112.10752","target":"arxiv:2204.03458","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2112.10752","target":"arxiv:2311.15127","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2112.10752","target":"arxiv:2311.17042","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2509.14252","target":"arxiv:2502.21321","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2509.14252","target":"arxiv:2403.08763","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2509.14252","target":"arxiv:2401.02954","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2509.14252","target":"arxiv:2407.12036","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1903.00374","target":"arxiv:2010.02193","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:1903.00374","target":"arxiv:2111.00210","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:1903.00374","target":"arxiv:2310.09615","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1903.00374","target":"arxiv:2007.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2204.01691","target":"arxiv:2305.10626","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2204.01691","target":"arxiv:2606.06556","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2204.01691","target":"arxiv:2511.22904","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2204.01691","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2404.08471","target":"arxiv:2506.09985","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2404.08471","target":"arxiv:2403.00504","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2404.08471","target":"arxiv:2307.12698","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2404.08471","target":"arxiv:2301.08243","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2509.12387","target":"arxiv:2102.11107","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2509.12387","target":"arxiv:2506.23068","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2509.12387","target":"arxiv:2505.02074","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2509.12387","target":"arxiv:2510.01539","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2508.12027","target":"arxiv:1909.10863","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2508.12027","target":"arxiv:2001.11841","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2508.12027","target":"arxiv:2603.20927","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2508.12027","target":"arxiv:1709.02341","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2208.06726","target":"arxiv:1701.06970","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2208.06726","target":"arxiv:2505.24631","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2208.06726","target":"arxiv:1510.00056","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2208.06726","target":"arxiv:1011.0063","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2303.08302","target":"arxiv:2305.17888","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2303.08302","target":"arxiv:2504.05352","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2303.08302","target":"arxiv:2406.06385","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2303.08302","target":"arxiv:2211.10438","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2202.03771","target":"arxiv:2203.03417","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2202.03771","target":"arxiv:2306.10754","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2202.03771","target":"arxiv:2212.13472","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2202.03771","target":"arxiv:2112.07103","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2301.04104","target":"nature.com:f005fae15467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.96},{"source":"arxiv:2301.04104","target":"arxiv:2010.02193","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2301.04104","target":"arxiv:2405.12399","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2301.04104","target":"arxiv:1803.10122","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2510.13161","target":"arxiv:2511.20340","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2510.13161","target":"arxiv:2603.03251","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2510.13161","target":"arxiv:2411.13157","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2510.13161","target":"arxiv:2402.01528","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2401.03804","target":"arxiv:2401.02954","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.76},{"source":"arxiv:2401.03804","target":"arxiv:2604.06882","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.74},{"source":"arxiv:2401.03804","target":"arxiv:2407.12036","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.73},{"source":"arxiv:2401.03804","target":"arxiv:2309.13638","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.73},{"source":"arxiv:2411.06284","target":"arxiv:2402.12451","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2411.06284","target":"arxiv:2404.07214","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2411.06284","target":"arxiv:2501.02189","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2411.06284","target":"arxiv:2508.13073","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2405.15383","target":"arxiv:2402.12275","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2405.15383","target":"arxiv:2305.14992","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2405.15383","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2405.15383","target":"arxiv:2502.13092","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2506.06725","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2506.06725","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2506.06725","target":"arxiv:2305.14992","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2506.06725","target":"arxiv:2605.23972","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2406.09455","target":"arxiv:2601.21998","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2406.09455","target":"arxiv:2603.28489","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2406.09455","target":"arxiv:2602.10717","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2406.09455","target":"arxiv:2511.08585","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"lumalabs.ai:cf7b765e8f15","target":"arxiv:2204.00227","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"lumalabs.ai:cf7b765e8f15","target":"arxiv:2105.05233","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"lumalabs.ai:cf7b765e8f15","target":"arxiv:2303.01469","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"lumalabs.ai:cf7b765e8f15","target":"arxiv:2406.07507","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"nature.com:386c3df42560","target":"arxiv:2312.15796","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"nature.com:386c3df42560","target":"arxiv:2212.12794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"nature.com:386c3df42560","target":"arxiv:2306.14066","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"nature.com:386c3df42560","target":"arxiv:2405.05925","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2006.04182","target":"arxiv:2508.08762","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2006.04182","target":"arxiv:2202.09467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2006.04182","target":"arxiv:2103.03725","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2006.04182","target":"arxiv:2308.07870","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2103.04723","target":"arxiv:2112.07103","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2103.04723","target":"arxiv:2001.10371","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2103.04723","target":"arxiv:2301.08861","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2103.04723","target":"arxiv:2212.13472","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2305.18153","target":"arxiv:2403.04121","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2305.18153","target":"arxiv:2506.21521","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2305.18153","target":"arxiv:2005.14165","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2305.18153","target":"arxiv:2205.11916","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2403.00504","target":"arxiv:2301.08243","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2403.00504","target":"arxiv:2404.08471","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2403.00504","target":"arxiv:2211.10831","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2403.00504","target":"arxiv:2602.22010","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"huggingface.co:9f24e732f5fe","target":"arxiv:2606.18250","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"huggingface.co:9f24e732f5fe","target":"arxiv:2605.12957","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"huggingface.co:9f24e732f5fe","target":"arxiv:2603.28489","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"huggingface.co:9f24e732f5fe","target":"arxiv:2605.06192","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2305.07017","target":"arxiv:2110.04544","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2305.07017","target":"arxiv:2406.06973","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2305.07017","target":"arxiv:2109.01134","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2305.07017","target":"arxiv:2103.00020","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"doi.org:0a734c9c22c9","target":"arxiv:1706.03741","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"doi.org:0a734c9c22c9","target":"arxiv:2208.01769","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"doi.org:0a734c9c22c9","target":"arxiv:1708.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"doi.org:0a734c9c22c9","target":"arxiv:1709.06560","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2406.11944","target":"arxiv:2309.08600","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2406.11944","target":"arxiv:2410.11767","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2406.11944","target":"transformer-circuits.pub:e2306116790e","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2406.11944","target":"arxiv:2202.08906","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2206.11795","target":"arxiv:2310.06114","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2206.11795","target":"arxiv:2602.10717","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2206.11795","target":"arxiv:2506.09985","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2206.11795","target":"arxiv:1903.00374","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2111.00210","target":"arxiv:1903.00374","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2111.00210","target":"arxiv:2010.02193","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2111.00210","target":"arxiv:1911.08265","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2111.00210","target":"arxiv:1706.03741","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1709.02341","target":"arxiv:2001.11841","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:1709.02341","target":"arxiv:2505.19867","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:1709.02341","target":"arxiv:1904.08149","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:1709.02341","target":"arxiv:1911.10601","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2311.16038","target":"arxiv:2502.10498","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2311.16038","target":"arxiv:2403.02622","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2311.16038","target":"arxiv:2309.09777","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2311.16038","target":"arxiv:2311.01017","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2412.18106","target":"arxiv:2504.09590","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2412.18106","target":"arxiv:2504.19516","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2412.18106","target":"arxiv:2503.20552","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2412.18106","target":"arxiv:2308.09723","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2306.14066","target":"arxiv:2312.15796","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2306.14066","target":"nature.com:386c3df42560","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2306.14066","target":"arxiv:2406.14798","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2306.14066","target":"arxiv:2405.05925","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2505.09694","target":"arxiv:2510.16732","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2505.09694","target":"arxiv:2605.06192","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2505.09694","target":"arxiv:2502.20694","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2505.09694","target":"arxiv:2410.18072","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2308.16369","target":"arxiv:2503.20552","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2308.16369","target":"arxiv:2504.19516","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2308.16369","target":"arxiv:2511.20340","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2308.16369","target":"arxiv:2401.07851","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2411.00769","target":"arxiv:2412.00887","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2411.00769","target":"arxiv:2605.12957","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2411.00769","target":"arxiv:2408.14837","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2411.00769","target":"arxiv:2506.18701","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2305.05383","target":"arxiv:2305.11169","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2305.05383","target":"arxiv:2402.09664","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2305.05383","target":"arxiv:2405.15383","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2305.05383","target":"arxiv:2503.02854","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2509.19958","target":"arxiv:2507.12440","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2509.19958","target":"arxiv:2606.06556","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2509.19958","target":"arxiv:2603.18532","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2509.19958","target":"arxiv:2508.13073","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2605.16547","target":"arxiv:2505.01712","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2605.16547","target":"arxiv:2503.15168","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2605.16547","target":"arxiv:2601.21998","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2605.16547","target":"arxiv:2604.06882","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2501.03575","target":"blogs.nvidia.com:9f8a6e618fbd","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"arxiv:2501.03575","target":"huggingface.co:e9042540da12","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2501.03575","target":"arxiv:2511.00062","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2501.03575","target":"github.com:14ee211b0b5d","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2411.10171","target":"arxiv:2403.02622","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2411.10171","target":"arxiv:2210.07729","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2411.10171","target":"arxiv:2506.18897","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2411.10171","target":"arxiv:2405.17398","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"nianticlabs.com:05a236f27efe","target":"reply.com:5748177416db","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"nianticlabs.com:05a236f27efe","target":"arxiv:2511.15722","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"nianticlabs.com:05a236f27efe","target":"arxiv:2510.25760","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"nianticlabs.com:05a236f27efe","target":"thesequence.substack.com:0f25bda15c4b","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2407.06886","target":"arxiv:2509.20021","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2407.06886","target":"arxiv:2507.00917","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2407.06886","target":"arxiv:2510.16732","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2407.06886","target":"arxiv:2506.22355","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2510.16732","target":"fi.ee.tsinghua.edu.cn:1c1fbee23e33","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2510.16732","target":"arxiv:2507.00917","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2510.16732","target":"arxiv:2605.12090","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2510.16732","target":"arxiv:2506.22355","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:1810.04805","target":"arxiv:1910.10683","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:1810.04805","target":"nature.com:a4d8ac562091","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:1810.04805","target":"arxiv:2402.10644","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:1810.04805","target":"arxiv:1706.03762","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2402.16823","target":"arxiv:2309.07870","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2402.16823","target":"arxiv:2502.12130","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2402.16823","target":"arxiv:2606.15874","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2402.16823","target":"arxiv:2402.02716","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"nature.com:e8ee4994d788","target":"nature.com:a4d8ac562091","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"nature.com:e8ee4994d788","target":"arxiv:2308.07870","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"nature.com:e8ee4994d788","target":"arxiv:2512.02419","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"nature.com:e8ee4994d788","target":"arxiv:2106.00737","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1911.08265","target":"arxiv:2111.00210","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1911.08265","target":"nature.com:e56ca063b728","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1911.08265","target":"arxiv:1903.00374","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1911.08265","target":"arxiv:2010.02193","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2307.15054","target":"arxiv:2311.03658","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2307.15054","target":"arxiv:2406.01506","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2307.15054","target":"transformer-circuits.pub:c4d9a1f0e2b7","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2307.15054","target":"arxiv:2507.21141","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2006.11239","target":"arxiv:2204.00227","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2006.11239","target":"arxiv:2105.05233","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2006.11239","target":"arxiv:2112.10752","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2006.11239","target":"arxiv:1907.05600","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2408.03943","target":"arxiv:2503.15168","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2408.03943","target":"arxiv:2604.16592","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2408.03943","target":"openreview.net:3f48a33ff9bf","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2408.03943","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2403.16971","target":"arxiv:2312.03815","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2403.16971","target":"arxiv:2601.01885","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2403.16971","target":"arxiv:2308.03688","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2403.16971","target":"arxiv:2309.07870","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2504.10612","target":"arxiv:2210.02747","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2504.10612","target":"arxiv:2304.14772","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2504.10612","target":"arxiv:2410.02548","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2504.10612","target":"arxiv:2406.07507","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2105.11203","target":"arxiv:2108.13343","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2105.11203","target":"arxiv:2002.04501","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2105.11203","target":"arxiv:2201.06387","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2105.11203","target":"arxiv:1906.10184","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2602.23152","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2602.23152","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2602.23152","target":"arxiv:2503.02143","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2602.23152","target":"arxiv:2407.00118","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2506.08967","target":"arxiv:2502.11946","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2506.08967","target":"arxiv:2507.02768","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2506.08967","target":"arxiv:2304.12995","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2506.08967","target":"arxiv:2405.08295","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"the-decoder.com:97885f462d1f","target":"arxiv:2606.09032","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"the-decoder.com:97885f462d1f","target":"arxiv:2305.10626","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"the-decoder.com:97885f462d1f","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"the-decoder.com:97885f462d1f","target":"arxiv:2406.06485","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2410.13056","target":"arxiv:2305.17888","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2410.13056","target":"arxiv:2504.05352","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2410.13056","target":"arxiv:2211.10438","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2410.13056","target":"arxiv:2312.03788","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2504.10903","target":"arxiv:2503.16419","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2504.10903","target":"arxiv:2502.18080","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2504.10903","target":"arxiv:2507.02076","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2504.10903","target":"arxiv:2502.03373","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2107.08241","target":"arxiv:2102.08363","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2107.08241","target":"arxiv:1709.06560","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2107.08241","target":"arxiv:2106.02039","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2107.08241","target":"arxiv:2502.14819","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2507.06952","target":"arxiv:2510.19788","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2507.06952","target":"arxiv:2511.12239","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2507.06952","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2507.06952","target":"arxiv:2406.03689","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2404.07214","target":"arxiv:2402.12451","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2404.07214","target":"arxiv:2411.06284","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2404.07214","target":"arxiv:2501.02189","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2404.07214","target":"arxiv:2306.09265","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2510.07092","target":"arxiv:2510.16732","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2510.07092","target":"arxiv:2606.18250","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2510.07092","target":"arxiv:2401.09985","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2510.07092","target":"arxiv:2511.08585","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2606.00133","target":"arxiv:2411.14499","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2606.00133","target":"reply.com:5748177416db","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2606.00133","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2606.00133","target":"arxiv:2507.05169","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2102.08363","target":"arxiv:2006.15134","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2102.08363","target":"arxiv:2107.08241","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2102.08363","target":"arxiv:2404.16767","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2102.08363","target":"arxiv:2106.02039","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2111.09259","target":"arxiv:2406.00877","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2111.09259","target":"arxiv:2403.15498","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2111.09259","target":"arxiv:1911.08265","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2111.09259","target":"arxiv:2309.00941","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2209.15571","target":"arxiv:2406.07507","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2209.15571","target":"arxiv:2210.02747","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2209.15571","target":"arxiv:2304.14772","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2209.15571","target":"arxiv:2410.02548","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2402.03161","target":"arxiv:2312.14125","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2402.03161","target":"arxiv:2602.08024","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2402.03161","target":"arxiv:2305.18264","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2402.03161","target":"arxiv:2406.06973","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2605.06192","target":"arxiv:2605.12090","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2605.06192","target":"arxiv:2601.21998","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2605.06192","target":"arxiv:2506.18897","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2605.06192","target":"arxiv:2602.10717","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2311.15127","target":"arxiv:2204.03458","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2311.15127","target":"arxiv:2112.10752","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2311.15127","target":"arxiv:2311.17042","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2311.15127","target":"arxiv:2303.01469","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2506.09849","target":"arxiv:1803.07616","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2506.09849","target":"arxiv:2502.11831","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2506.09849","target":"arxiv:2501.09038","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2506.09849","target":"arxiv:2503.06800","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2410.24164","target":"arxiv:2502.05855","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2410.24164","target":"arxiv:2307.15818","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2410.24164","target":"arxiv:2510.17111","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2410.24164","target":"arxiv:2508.13073","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2502.13092","target":"arxiv:2305.14909","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2502.13092","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2502.13092","target":"arxiv:2606.09032","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2502.13092","target":"arxiv:2305.14992","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2505.24631","target":"arxiv:1510.00056","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2505.24631","target":"arxiv:2208.06726","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2505.24631","target":"arxiv:1701.06970","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2505.24631","target":"arxiv:0910.3570","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.76},{"source":"arxiv:2106.00737","target":"arxiv:2305.11169","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2106.00737","target":"arxiv:2210.13382","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2106.00737","target":"nature.com:a4d8ac562091","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2106.00737","target":"nature.com:e8ee4994d788","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2310.16828","target":"arxiv:2107.08241","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2310.16828","target":"arxiv:2411.10171","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2310.16828","target":"arxiv:2010.02193","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2310.16828","target":"arxiv:2406.15836","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:1801.04819","target":"arxiv:2507.00917","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:1801.04819","target":"arxiv:2407.06886","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1801.04819","target":"arxiv:2509.20021","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1801.04819","target":"arxiv:2408.03943","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2605.16395","target":"arxiv:2310.06114","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2605.16395","target":"arxiv:2507.00917","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2605.16395","target":"reply.com:5748177416db","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2605.16395","target":"arxiv:2510.16732","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"nature.com:a4d8ac562091","target":"nature.com:e8ee4994d788","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"nature.com:a4d8ac562091","target":"arxiv:2512.02419","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"nature.com:a4d8ac562091","target":"arxiv:2308.07870","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"nature.com:a4d8ac562091","target":"arxiv:2406.16062","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2507.12440","target":"arxiv:2502.05855","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2507.12440","target":"arxiv:2508.13073","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2507.12440","target":"arxiv:2509.19958","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2507.12440","target":"arxiv:2510.17111","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"openaipublic.blob.core.windows.net:570a1f9c6e10","target":"nature.com:a4d8ac562091","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"openaipublic.blob.core.windows.net:570a1f9c6e10","target":"nature.com:e8ee4994d788","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"openaipublic.blob.core.windows.net:570a1f9c6e10","target":"arxiv:2309.08600","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"openaipublic.blob.core.windows.net:570a1f9c6e10","target":"transformer-circuits.pub:b904370d3db5","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2306.05685","target":"arxiv:2411.15594","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2306.05685","target":"arxiv:2305.17926","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2306.05685","target":"arxiv:2507.10535","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2306.05685","target":"arxiv:2307.16125","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2407.12036","target":"arxiv:2502.21321","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2407.12036","target":"arxiv:2405.06211","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2407.12036","target":"arxiv:2403.04121","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2407.12036","target":"arxiv:2401.02954","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2305.14992","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2305.14992","target":"arxiv:2403.04121","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2305.14992","target":"arxiv:2305.14909","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2305.14992","target":"arxiv:2309.17382","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2407.10311","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2407.10311","target":"arxiv:2601.17067","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2407.10311","target":"x.com:351692e95fe5","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2407.10311","target":"arxiv:2405.03520","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2411.04983","target":"arxiv:2602.22010","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2411.04983","target":"arxiv:2510.19818","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2411.04983","target":"arxiv:2010.02193","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2411.04983","target":"arxiv:2605.06192","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2606.06556","target":"pi.website:ca5c1a691358","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2606.06556","target":"arxiv:2509.19958","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2606.06556","target":"arxiv:2603.18532","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2606.06556","target":"arxiv:2511.02097","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2603.13910","target":"arxiv:2506.01799","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2603.13910","target":"arxiv:2605.12957","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2603.13910","target":"huggingface.co:9f24e732f5fe","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2603.13910","target":"arxiv:2406.07472","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2509.02547","target":"arxiv:2208.01769","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2509.02547","target":"arxiv:2601.01885","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2509.02547","target":"arxiv:2603.07670","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2509.02547","target":"arxiv:2007.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2501.02189","target":"arxiv:2411.06284","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2501.02189","target":"arxiv:2404.07214","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2501.02189","target":"arxiv:2402.12451","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2501.02189","target":"arxiv:2306.09265","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2509.13281","target":"arxiv:2507.21141","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2509.13281","target":"arxiv:2407.15549","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2509.13281","target":"arxiv:2501.18837","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2509.13281","target":"arxiv:2307.15054","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2511.20340","target":"arxiv:2411.13157","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2511.20340","target":"arxiv:2401.07851","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2511.20340","target":"arxiv:2510.13161","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2511.20340","target":"arxiv:2603.03251","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"github.com:14ee211b0b5d","target":"blogs.nvidia.com:9f8a6e618fbd","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"github.com:14ee211b0b5d","target":"arxiv:2501.03575","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"github.com:14ee211b0b5d","target":"huggingface.co:e9042540da12","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"github.com:14ee211b0b5d","target":"arxiv:2511.00062","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"people.idsia.ch:73d35fac50f9","target":"arxiv:1803.10122","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"people.idsia.ch:73d35fac50f9","target":"arxiv:1706.03741","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"people.idsia.ch:73d35fac50f9","target":"arxiv:2509.04633","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"people.idsia.ch:73d35fac50f9","target":"openreview.net:3f48a33ff9bf","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2506.21521","target":"arxiv:2403.04121","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2506.21521","target":"arxiv:2305.18153","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2506.21521","target":"arxiv:2305.15074","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2506.21521","target":"arxiv:2309.13638","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2411.08794","target":"arxiv:2305.14992","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2411.08794","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2411.08794","target":"arxiv:2403.04121","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2411.08794","target":"arxiv:2605.23972","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2502.14739","target":"arxiv:2307.13692","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2502.14739","target":"arxiv:2305.15074","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2502.14739","target":"arxiv:2307.16125","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2502.14739","target":"arxiv:2401.02954","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2101.07393","target":"arxiv:1708.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2101.07393","target":"arxiv:1910.08210","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2101.07393","target":"arxiv:2511.22904","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2101.07393","target":"arxiv:2402.01695","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2304.03442","target":"arxiv:2410.02664","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2304.03442","target":"arxiv:2506.22355","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2304.03442","target":"arxiv:2402.15391","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2304.03442","target":"arxiv:2210.13382","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2412.07077","target":"arxiv:2109.01134","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2412.07077","target":"arxiv:2110.04544","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2412.07077","target":"arxiv:2406.06973","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2412.07077","target":"arxiv:2103.00020","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2503.15558","target":"huggingface.co:e9042540da12","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2503.15558","target":"arxiv:2501.03575","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2503.15558","target":"arxiv:2510.04978","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2503.15558","target":"blogs.nvidia.com:9f8a6e618fbd","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"pubmed.ncbi.nlm.nih.gov:fd66789ead66","target":"nature.com:ef956cce5b54","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"pubmed.ncbi.nlm.nih.gov:fd66789ead66","target":"nature.com:c3aa5b0d9273","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"pubmed.ncbi.nlm.nih.gov:fd66789ead66","target":"nature.com:ad56347f1921","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"pubmed.ncbi.nlm.nih.gov:fd66789ead66","target":"arxiv:1712.04602","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2502.10498","target":"arxiv:2403.02622","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2502.10498","target":"arxiv:2309.09777","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2502.10498","target":"arxiv:2311.17918","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2502.10498","target":"arxiv:2311.13549","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"nature.com:ad56347f1921","target":"arxiv:2402.06590","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"nature.com:ad56347f1921","target":"arxiv:2202.01682","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"nature.com:ad56347f1921","target":"nature.com:ef956cce5b54","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"nature.com:ad56347f1921","target":"nature.com:c3aa5b0d9273","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2405.17398","target":"arxiv:2411.10171","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2405.17398","target":"arxiv:2502.10498","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2405.17398","target":"arxiv:2403.02622","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2405.17398","target":"arxiv:2309.17080","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2507.21141","target":"arxiv:2509.13281","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2507.21141","target":"arxiv:2407.15549","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2507.21141","target":"arxiv:2603.27918","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2507.21141","target":"arxiv:2307.15054","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2511.20644","target":"arxiv:2510.08531","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2511.20644","target":"arxiv:2509.06266","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2511.20644","target":"arxiv:2602.03916","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2511.20644","target":"arxiv:2510.25760","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2310.04276","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2310.04276","target":"arxiv:2503.15168","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2310.04276","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2310.04276","target":"arxiv:2411.14499","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2511.08585","target":"arxiv:2603.28489","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2511.08585","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2511.08585","target":"arxiv:2601.17067","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2511.08585","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2205.11916","target":"arxiv:2201.11903","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2205.11916","target":"arxiv:2503.16419","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2205.11916","target":"arxiv:2502.03373","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2205.11916","target":"arxiv:2305.14992","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2605.09131","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2605.09131","target":"arxiv:2302.01560","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2605.09131","target":"arxiv:2305.10626","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2605.09131","target":"arxiv:2606.09032","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2406.18043","target":"arxiv:2603.18532","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2406.18043","target":"arxiv:2308.01399","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2406.18043","target":"arxiv:2410.02664","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2406.18043","target":"arxiv:2508.13073","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2303.00915","target":"arxiv:2506.22567","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2303.00915","target":"arxiv:2210.04133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2303.00915","target":"arxiv:2507.05201","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2303.00915","target":"arxiv:2512.00597","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"philpapers.org:9a9c4dcb81d9","target":"philpapers.org:4f7ed9918446","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"philpapers.org:9a9c4dcb81d9","target":"philpapers.org:7b684275d20a","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"philpapers.org:9a9c4dcb81d9","target":"philpapers.org:e7fc6fd114b8","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"philpapers.org:9a9c4dcb81d9","target":"philpapers.org:8c0576aa7f12","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:1701.06970","target":"arxiv:1510.00056","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:1701.06970","target":"arxiv:2208.06726","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1701.06970","target":"arxiv:2505.24631","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:1701.06970","target":"arxiv:1911.03537","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.75},{"source":"arxiv:2411.02385","target":"arxiv:2502.20694","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2411.02385","target":"arxiv:2501.09038","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2411.02385","target":"arxiv:2603.28489","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2411.02385","target":"arxiv:2406.03520","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"blogs.nvidia.com:9f8a6e618fbd","target":"arxiv:2501.03575","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"blogs.nvidia.com:9f8a6e618fbd","target":"huggingface.co:e9042540da12","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.94},{"source":"blogs.nvidia.com:9f8a6e618fbd","target":"arxiv:2511.00062","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"blogs.nvidia.com:9f8a6e618fbd","target":"github.com:14ee211b0b5d","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2506.22355","target":"arxiv:2509.20021","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2506.22355","target":"arxiv:2507.00917","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2506.22355","target":"arxiv:2510.16732","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2506.22355","target":"arxiv:2407.06886","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2402.10196","target":"arxiv:2603.27918","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2402.10196","target":"arxiv:2605.17288","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2402.10196","target":"arxiv:2606.10749","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2402.10196","target":"arxiv:2407.15549","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2304.14772","target":"arxiv:2210.02747","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2304.14772","target":"arxiv:2410.02548","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2304.14772","target":"arxiv:2406.07507","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2304.14772","target":"arxiv:2504.10612","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2512.21577","target":"arxiv:2511.12239","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2512.21577","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2512.21577","target":"arxiv:2407.10311","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2512.21577","target":"arxiv:2310.04276","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.76},{"source":"arxiv:2603.20224","target":"arxiv:2507.02076","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2603.20224","target":"arxiv:2502.18080","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2603.20224","target":"arxiv:2510.03847","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2603.20224","target":"arxiv:2511.20340","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2302.01560","target":"arxiv:2402.02716","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2302.01560","target":"arxiv:2305.14909","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2302.01560","target":"arxiv:2305.16653","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2302.01560","target":"arxiv:2605.09131","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2311.01017","target":"openreview.net:c15700430713","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2311.01017","target":"arxiv:2411.10171","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2311.01017","target":"arxiv:2601.21998","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2311.01017","target":"arxiv:2412.03572","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1904.08149","target":"arxiv:2001.11841","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:1904.08149","target":"arxiv:1911.10601","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:1904.08149","target":"arxiv:1909.10863","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:1904.08149","target":"arxiv:1709.02341","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2507.02076","target":"arxiv:2502.18080","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2507.02076","target":"arxiv:2503.16419","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2507.02076","target":"arxiv:2504.10903","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2507.02076","target":"arxiv:2603.20224","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"engraved.blog:b8c16edd622e","target":"arxiv:2304.12995","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"engraved.blog:b8c16edd622e","target":"arxiv:2310.08560","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.73},{"source":"engraved.blog:b8c16edd622e","target":"arxiv:2406.09455","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.72},{"source":"engraved.blog:b8c16edd622e","target":"arxiv:2402.15391","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.71},{"source":"neurips.cc:9c7e4f09bb35","target":"arxiv:2412.14161","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.96},{"source":"neurips.cc:9c7e4f09bb35","target":"arxiv:2308.03688","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"neurips.cc:9c7e4f09bb35","target":"arxiv:2503.16416","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"neurips.cc:9c7e4f09bb35","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"huggingface.co:e9042540da12","target":"blogs.nvidia.com:9f8a6e618fbd","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.94},{"source":"huggingface.co:e9042540da12","target":"arxiv:2511.00062","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.94},{"source":"huggingface.co:e9042540da12","target":"arxiv:2501.03575","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"huggingface.co:e9042540da12","target":"github.com:14ee211b0b5d","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2503.20523","target":"arxiv:2309.17080","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"arxiv:2503.20523","target":"arxiv:2405.17398","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2503.20523","target":"arxiv:2411.10171","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2503.20523","target":"arxiv:2403.02622","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2605.07234","target":"arxiv:2603.20397","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2605.07234","target":"arxiv:2412.02252","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2605.07234","target":"arxiv:2606.24467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2605.07234","target":"arxiv:2508.06297","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:1910.08210","target":"arxiv:2101.07393","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1910.08210","target":"arxiv:2511.22904","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1910.08210","target":"arxiv:1708.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1910.08210","target":"arxiv:2406.18043","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2309.12288","target":"arxiv:2309.13638","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2309.12288","target":"arxiv:2502.21321","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2309.12288","target":"arxiv:2305.18153","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2309.12288","target":"arxiv:2502.03373","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2412.02252","target":"arxiv:2606.24467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2412.02252","target":"arxiv:2605.07234","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2412.02252","target":"arxiv:2508.06297","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2412.02252","target":"arxiv:2410.00161","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2407.01392","target":"arxiv:2405.12399","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2407.01392","target":"arxiv:2204.00227","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2407.01392","target":"arxiv:2204.03458","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2407.01392","target":"arxiv:2006.11239","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2503.20552","target":"arxiv:2504.19516","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2503.20552","target":"arxiv:2308.16369","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2503.20552","target":"arxiv:2412.18106","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2503.20552","target":"arxiv:2504.09590","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:1706.02275","target":"arxiv:1812.00922","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:1706.02275","target":"arxiv:2208.01769","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:1706.02275","target":"arxiv:1711.00832","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1706.02275","target":"arxiv:2406.15836","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"philpapers.org:4f7ed9918446","target":"philpapers.org:9a9c4dcb81d9","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"philpapers.org:4f7ed9918446","target":"philpapers.org:7b684275d20a","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"philpapers.org:4f7ed9918446","target":"philpapers.org:e7fc6fd114b8","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"philpapers.org:4f7ed9918446","target":"philpapers.org:8c0576aa7f12","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2307.13692","target":"arxiv:2305.15074","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2307.13692","target":"arxiv:2502.14739","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2307.13692","target":"arxiv:2307.16125","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2307.13692","target":"arxiv:2309.13638","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2308.09723","target":"arxiv:2211.10438","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2308.09723","target":"arxiv:2312.03788","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2308.09723","target":"arxiv:2305.17888","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2308.09723","target":"arxiv:2406.06385","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2509.00579","target":"arxiv:2410.00161","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2509.00579","target":"arxiv:2603.20397","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2509.00579","target":"arxiv:2606.24467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2509.00579","target":"arxiv:2508.06297","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2310.03744","target":"arxiv:2402.12451","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2310.03744","target":"arxiv:2311.05348","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2310.03744","target":"arxiv:2306.09265","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2310.03744","target":"arxiv:2501.02189","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1812.00922","target":"arxiv:1706.02275","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:1812.00922","target":"arxiv:2208.01769","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:1812.00922","target":"arxiv:1711.00832","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1812.00922","target":"arxiv:2305.17198","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2305.18654","target":"arxiv:2508.16745","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2305.18654","target":"arxiv:2502.18080","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2305.18654","target":"arxiv:2309.13638","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2305.18654","target":"arxiv:2407.12036","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2301.08243","target":"arxiv:2403.00504","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2301.08243","target":"arxiv:2307.12698","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2301.08243","target":"arxiv:2211.10831","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2301.08243","target":"arxiv:2404.08471","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"philpapers.org:8c0576aa7f12","target":"philpapers.org:f2d0d9f7c8fe","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"philpapers.org:8c0576aa7f12","target":"philpapers.org:e7fc6fd114b8","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"philpapers.org:8c0576aa7f12","target":"philpapers.org:7b684275d20a","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"philpapers.org:8c0576aa7f12","target":"philpapers.org:9a9c4dcb81d9","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2406.03689","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2406.03689","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2406.03689","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2406.03689","target":"arxiv:2210.13382","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2405.03520","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2405.03520","target":"arxiv:2511.08585","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2405.03520","target":"arxiv:2411.14499","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2405.03520","target":"technologyreview.com:6e6b17c22c9b","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"introl.com:c695c416f507","target":"thesequence.substack.com:0f25bda15c4b","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"introl.com:c695c416f507","target":"technologyreview.com:6e6b17c22c9b","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"introl.com:c695c416f507","target":"blogs.nvidia.com:9f8a6e618fbd","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"introl.com:c695c416f507","target":"github.com:14ee211b0b5d","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2603.13647","target":"arxiv:2605.16689","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2603.13647","target":"arxiv:2604.06882","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.76},{"source":"arxiv:2603.13647","target":"arxiv:2410.10377","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.75},{"source":"arxiv:2603.13647","target":"arxiv:2505.11329","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.74},{"source":"arxiv:2410.05229","target":"arxiv:2205.11916","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2410.05229","target":"arxiv:2309.13638","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2410.05229","target":"arxiv:2307.13692","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2410.05229","target":"arxiv:2305.14992","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2405.05925","target":"arxiv:2312.15796","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2405.05925","target":"nature.com:386c3df42560","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2405.05925","target":"arxiv:2306.14066","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2405.05925","target":"arxiv:2212.12794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2405.05967","target":"arxiv:2311.17042","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2405.05967","target":"arxiv:2406.02347","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2405.05967","target":"arxiv:2105.05233","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2405.05967","target":"arxiv:2006.11239","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2411.14499","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2411.14499","target":"technologyreview.com:6e6b17c22c9b","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2411.14499","target":"arxiv:2511.02097","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2411.14499","target":"arxiv:2510.21219","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2304.09116","target":"arxiv:2102.01192","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2304.09116","target":"arxiv:2405.08295","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2304.09116","target":"arxiv:2506.08967","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2304.09116","target":"arxiv:2502.11946","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2403.04253","target":"arxiv:2010.02193","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2403.04253","target":"arxiv:1803.10122","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2403.04253","target":"nature.com:f005fae15467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2403.04253","target":"arxiv:2301.04104","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2606.18250","target":"huggingface.co:9f24e732f5fe","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2606.18250","target":"arxiv:2605.06192","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2606.18250","target":"arxiv:2511.08585","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2606.18250","target":"arxiv:2406.07472","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"pubmed.ncbi.nlm.nih.gov:9eeb70dd85f3","target":"arxiv:2402.06590","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"pubmed.ncbi.nlm.nih.gov:9eeb70dd85f3","target":"pubmed.ncbi.nlm.nih.gov:fd66789ead66","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"pubmed.ncbi.nlm.nih.gov:9eeb70dd85f3","target":"nature.com:a4d8ac562091","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"pubmed.ncbi.nlm.nih.gov:9eeb70dd85f3","target":"arxiv:2107.08241","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2303.11366","target":"arxiv:2509.02547","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2303.11366","target":"arxiv:2210.03629","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2303.11366","target":"arxiv:2502.12130","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2303.11366","target":"arxiv:2308.01399","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2503.16416","target":"arxiv:2308.03688","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2503.16416","target":"arxiv:2401.13178","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2503.16416","target":"arxiv:2412.14161","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2503.16416","target":"arxiv:2606.10749","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2108.13343","target":"arxiv:2201.06387","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.94},{"source":"arxiv:2108.13343","target":"arxiv:2105.11203","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2108.13343","target":"arxiv:1705.09156","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2108.13343","target":"arxiv:2002.04501","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2212.06817","target":"arxiv:2307.15818","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2212.06817","target":"arxiv:2310.08864","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2212.06817","target":"pi.website:ca5c1a691358","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2212.06817","target":"arxiv:2603.18532","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2505.24618","target":"arxiv:2603.20927","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2505.24618","target":"arxiv:2505.19867","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2505.24618","target":"arxiv:1909.10863","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2505.24618","target":"arxiv:1911.10601","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:1510.00056","target":"arxiv:2505.24631","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:1510.00056","target":"arxiv:1701.06970","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:1510.00056","target":"arxiv:2208.06726","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:1510.00056","target":"arxiv:0910.3570","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.75},{"source":"arxiv:2506.01799","target":"arxiv:2603.13910","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2506.01799","target":"arxiv:2401.09985","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2506.01799","target":"arxiv:2406.07472","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2506.01799","target":"arxiv:2605.12957","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2412.03572","target":"arxiv:2601.21998","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2412.03572","target":"arxiv:2403.02622","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2412.03572","target":"arxiv:2311.01017","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2412.03572","target":"arxiv:2602.22010","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2501.16411","target":"arxiv:2106.08261","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2501.16411","target":"arxiv:2510.04978","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2501.16411","target":"arxiv:2505.09694","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2501.16411","target":"arxiv:2510.17111","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2406.02347","target":"arxiv:2405.05967","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2406.02347","target":"arxiv:2311.17042","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2406.02347","target":"arxiv:2112.10752","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2406.02347","target":"arxiv:2204.03458","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2106.08261","target":"arxiv:2306.15668","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2106.08261","target":"arxiv:2501.16411","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2106.08261","target":"arxiv:2509.04633","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2106.08261","target":"arxiv:2510.04978","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2512.04513","target":"arxiv:2602.15400","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2512.04513","target":"arxiv:2305.10626","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2512.04513","target":"arxiv:2407.00118","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2512.04513","target":"arxiv:2402.12451","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2109.06780","target":"arxiv:2206.08853","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2109.06780","target":"arxiv:2308.03688","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2109.06780","target":"arxiv:2401.13178","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2109.06780","target":"arxiv:2509.02547","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2407.18003","target":"arxiv:2603.20397","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2407.18003","target":"arxiv:2410.00161","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2407.18003","target":"arxiv:2508.06297","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2407.18003","target":"arxiv:2606.24467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2503.20020","target":"pi.website:ca5c1a691358","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2503.20020","target":"arxiv:2508.13073","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2503.20020","target":"arxiv:2510.17111","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2503.20020","target":"arxiv:2507.12440","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2406.06973","target":"arxiv:2110.04544","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2406.06973","target":"arxiv:2109.01134","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2406.06973","target":"arxiv:2103.00020","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2406.06973","target":"arxiv:2412.07077","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2106.02039","target":"arxiv:2502.14819","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2106.02039","target":"arxiv:2107.08241","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2106.02039","target":"arxiv:2102.08363","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2106.02039","target":"arxiv:1706.03741","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2105.05233","target":"arxiv:2112.10752","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2105.05233","target":"arxiv:2311.17042","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2105.05233","target":"arxiv:2405.05967","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2105.05233","target":"arxiv:2204.03458","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2412.00887","target":"arxiv:2408.14837","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2412.00887","target":"arxiv:2411.00769","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2412.00887","target":"arxiv:2506.18701","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2412.00887","target":"arxiv:2605.12957","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2603.07670","target":"arxiv:2601.01885","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2603.07670","target":"arxiv:2402.02716","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2603.07670","target":"arxiv:2509.02547","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2603.07670","target":"arxiv:2503.16416","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2605.16689","target":"arxiv:2604.06882","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2605.16689","target":"arxiv:2605.16547","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2605.16689","target":"arxiv:2603.13647","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2605.16689","target":"arxiv:2505.24618","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2506.09042","target":"blogs.nvidia.com:9f8a6e618fbd","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2506.09042","target":"arxiv:2501.03575","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2506.09042","target":"huggingface.co:e9042540da12","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2506.09042","target":"arxiv:2511.00062","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2202.01682","target":"frontiersin.org:d1ec3c5c2c89","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2202.01682","target":"nature.com:ad56347f1921","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2202.01682","target":"nature.com:c3aa5b0d9273","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2202.01682","target":"arxiv:2105.07308","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2505.11329","target":"arxiv:2603.20224","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2505.11329","target":"arxiv:2504.09590","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2505.11329","target":"arxiv:2308.16369","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2505.11329","target":"arxiv:2412.18106","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"oasis-model.github.io:525053739707","target":"arxiv:2402.15391","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"oasis-model.github.io:525053739707","target":"arxiv:2209.00588","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"oasis-model.github.io:525053739707","target":"thesequence.substack.com:0f25bda15c4b","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.76},{"source":"oasis-model.github.io:525053739707","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.76},{"source":"arxiv:2006.15134","target":"arxiv:2102.08363","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2006.15134","target":"arxiv:2404.16767","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2006.15134","target":"arxiv:2007.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2006.15134","target":"arxiv:1709.06560","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2502.18080","target":"arxiv:2502.03373","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2502.18080","target":"arxiv:2504.10903","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2502.18080","target":"arxiv:2503.16419","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2502.18080","target":"arxiv:2507.02076","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2405.06211","target":"arxiv:2312.10997","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"arxiv:2405.06211","target":"arxiv:2005.11401","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2405.06211","target":"arxiv:2404.00610","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2405.06211","target":"arxiv:2505.14146","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:1805.09176","target":"arxiv:2202.01682","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:1805.09176","target":"frontiersin.org:d1ec3c5c2c89","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:1805.09176","target":"nature.com:ef956cce5b54","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:1805.09176","target":"arxiv:2402.06590","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2509.04633","target":"arxiv:2503.15168","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2509.04633","target":"arxiv:2512.02419","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2509.04633","target":"nature.com:ef956cce5b54","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2509.04633","target":"arxiv:2402.06590","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2404.16767","target":"arxiv:2006.15134","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2404.16767","target":"arxiv:2007.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2404.16767","target":"arxiv:2102.08363","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2404.16767","target":"arxiv:1709.06560","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"deepmind.google:f020e8dd5471","target":"deepmind.google:39914de6d3c9","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"deepmind.google:f020e8dd5471","target":"arxiv:2402.15391","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"deepmind.google:f020e8dd5471","target":"reply.com:5748177416db","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"deepmind.google:f020e8dd5471","target":"arxiv:2406.09455","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2407.06508","target":"arxiv:2210.04133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2407.06508","target":"arxiv:2303.00915","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2407.06508","target":"arxiv:2506.22567","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2407.06508","target":"arxiv:2406.16062","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2410.02664","target":"arxiv:2406.15836","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2410.02664","target":"arxiv:2602.12520","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2410.02664","target":"arxiv:2308.01399","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2410.02664","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2403.02622","target":"arxiv:2502.10498","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2403.02622","target":"arxiv:2511.02097","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2403.02622","target":"arxiv:2411.14499","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2403.02622","target":"arxiv:2309.09777","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2502.20694","target":"arxiv:2410.18072","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2502.20694","target":"arxiv:2505.09694","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2502.20694","target":"arxiv:2401.09985","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2502.20694","target":"arxiv:2411.02385","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1711.00832","target":"arxiv:2208.01769","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:1711.00832","target":"arxiv:2305.17198","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:1711.00832","target":"arxiv:1706.02275","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1711.00832","target":"arxiv:1812.00922","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2010.02193","target":"arxiv:1903.00374","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2010.02193","target":"arxiv:2405.12399","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2010.02193","target":"nature.com:f005fae15467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2010.02193","target":"arxiv:2301.04104","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2504.19516","target":"arxiv:2503.20552","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2504.19516","target":"arxiv:2504.09590","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2504.19516","target":"arxiv:2308.16369","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2504.19516","target":"arxiv:2602.24044","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2312.14125","target":"arxiv:2402.03161","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2312.14125","target":"arxiv:2509.20328","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2312.14125","target":"arxiv:2305.18264","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2312.14125","target":"arxiv:2510.04999","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2311.13549","target":"arxiv:2403.02622","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2311.13549","target":"arxiv:2502.10498","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2311.13549","target":"arxiv:2309.09777","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2311.13549","target":"arxiv:2411.10171","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2509.20328","target":"arxiv:2510.26802","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2509.20328","target":"arxiv:2312.14125","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2509.20328","target":"arxiv:2501.09038","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2509.20328","target":"arxiv:2603.28489","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"philarchive.org:69d8d2ad404e","target":"philarchive.org:04bcc5cbd425","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.94},{"source":"philarchive.org:69d8d2ad404e","target":"philpapers.org:7b684275d20a","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"philarchive.org:69d8d2ad404e","target":"philpapers.org:e7fc6fd114b8","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"philarchive.org:69d8d2ad404e","target":"philpapers.org:8c0576aa7f12","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"nature.com:413d00acfd43","target":"arxiv:1805.07526","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"nature.com:413d00acfd43","target":"arxiv:2207.12316","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"nature.com:413d00acfd43","target":"arxiv:2308.07870","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"nature.com:413d00acfd43","target":"arxiv:2006.04182","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2510.04978","target":"arxiv:2407.06886","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2510.04978","target":"arxiv:2605.29930","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2510.04978","target":"arxiv:2310.08803","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2510.04978","target":"arxiv:2509.20021","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2605.22882","target":"arxiv:2406.07472","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2605.22882","target":"arxiv:2605.12957","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2605.22882","target":"arxiv:2606.18250","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2605.22882","target":"arxiv:2605.06192","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2308.07870","target":"arxiv:2202.09467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2308.07870","target":"arxiv:2508.08762","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2308.07870","target":"arxiv:2006.04182","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2308.07870","target":"arxiv:2207.12316","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"philarchive.org:04bcc5cbd425","target":"philarchive.org:69d8d2ad404e","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.94},{"source":"philarchive.org:04bcc5cbd425","target":"philpapers.org:7b684275d20a","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"philarchive.org:04bcc5cbd425","target":"philpapers.org:e7fc6fd114b8","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"philarchive.org:04bcc5cbd425","target":"philpapers.org:8c0576aa7f12","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"x.com:351692e95fe5","target":"arxiv:2601.17067","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"x.com:351692e95fe5","target":"arxiv:2211.10831","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"x.com:351692e95fe5","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"x.com:351692e95fe5","target":"arxiv:2407.10311","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1012.5151","target":"arxiv:1011.0063","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"arxiv:1012.5151","target":"arxiv:0910.3570","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1012.5151","target":"arxiv:1911.03537","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:1012.5151","target":"arxiv:2208.06726","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.75},{"source":"arxiv:2402.09664","target":"arxiv:2507.10535","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2402.09664","target":"arxiv:2305.14992","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2402.09664","target":"arxiv:2403.04121","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2402.09664","target":"arxiv:2305.05383","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2504.05352","target":"arxiv:2303.08302","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2504.05352","target":"arxiv:2305.17888","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2504.05352","target":"arxiv:2211.10438","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2504.05352","target":"arxiv:2312.03788","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:1906.10184","target":"arxiv:2201.06387","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.94},{"source":"arxiv:1906.10184","target":"arxiv:2002.04501","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:1906.10184","target":"arxiv:2105.11203","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:1906.10184","target":"arxiv:2108.13343","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2210.02747","target":"arxiv:2304.14772","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2210.02747","target":"arxiv:2410.02548","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2210.02747","target":"arxiv:2406.07507","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2210.02747","target":"arxiv:2209.15571","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2403.04121","target":"arxiv:2305.14992","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2403.04121","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2403.04121","target":"arxiv:2502.21321","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2403.04121","target":"arxiv:2305.14909","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2105.07308","target":"arxiv:2308.07870","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2105.07308","target":"arxiv:2402.06590","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2105.07308","target":"arxiv:2202.01682","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2105.07308","target":"arxiv:2512.02419","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2603.18532","target":"arxiv:2508.13073","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2603.18532","target":"arxiv:2606.06556","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2603.18532","target":"arxiv:2510.17111","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2603.18532","target":"arxiv:2509.19958","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2402.01528","target":"arxiv:2411.13157","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2402.01528","target":"arxiv:2401.07851","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2402.01528","target":"arxiv:2603.03251","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2402.01528","target":"arxiv:2511.20340","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2501.18837","target":"arxiv:2407.15549","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2501.18837","target":"arxiv:2507.21141","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2501.18837","target":"arxiv:2509.13281","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2501.18837","target":"arxiv:2603.27918","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2109.01134","target":"arxiv:2110.04544","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2109.01134","target":"arxiv:2412.07077","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2109.01134","target":"arxiv:2103.00020","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2109.01134","target":"arxiv:2406.06973","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2506.21539","target":"arxiv:2605.12090","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2506.21539","target":"arxiv:2602.22010","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2506.21539","target":"arxiv:2510.17111","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2506.21539","target":"arxiv:2605.06192","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2510.19818","target":"arxiv:2605.12090","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2510.19818","target":"arxiv:2602.22010","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2510.19818","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2510.19818","target":"arxiv:2411.14499","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2411.13112","target":"arxiv:2509.06266","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2411.13112","target":"arxiv:2510.08531","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2411.13112","target":"arxiv:2511.15722","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2411.13112","target":"arxiv:2510.25760","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:1706.03762","target":"arxiv:2405.10480","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1706.03762","target":"arxiv:2402.10644","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:1706.03762","target":"transformer-circuits.pub:e2306116790e","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:1706.03762","target":"arxiv:2406.11944","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2307.15818","target":"arxiv:2508.13073","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2307.15818","target":"arxiv:2212.06817","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2307.15818","target":"arxiv:2510.17111","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2307.15818","target":"pi.website:ca5c1a691358","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2212.13472","target":"arxiv:2103.04723","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2212.13472","target":"arxiv:2202.03771","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2212.13472","target":"arxiv:2112.07103","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2212.13472","target":"arxiv:2301.08861","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"nature.com:e173b62796c2","target":"nature.com:ef956cce5b54","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"nature.com:e173b62796c2","target":"nature.com:c3aa5b0d9273","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"nature.com:e173b62796c2","target":"nature.com:ad56347f1921","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"nature.com:e173b62796c2","target":"arxiv:1805.09042","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2505.19867","target":"arxiv:1709.02341","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2505.19867","target":"arxiv:2001.11841","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2505.19867","target":"arxiv:1911.10601","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2505.19867","target":"arxiv:2603.20927","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"openreview.net:3f48a33ff9bf","target":"arxiv:2306.02572","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"openreview.net:3f48a33ff9bf","target":"arxiv:2503.15168","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"openreview.net:3f48a33ff9bf","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"openreview.net:3f48a33ff9bf","target":"arxiv:2402.06590","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2503.16419","target":"arxiv:2504.10903","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2503.16419","target":"arxiv:2502.18080","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2503.16419","target":"arxiv:2502.03373","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2503.16419","target":"arxiv:2507.02076","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2602.08024","target":"arxiv:2402.03161","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2602.08024","target":"arxiv:2311.05348","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2602.08024","target":"arxiv:2510.04999","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2602.08024","target":"arxiv:2312.14125","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2410.10377","target":"arxiv:2505.01712","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2410.10377","target":"arxiv:2007.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.76},{"source":"arxiv:2410.10377","target":"arxiv:2603.21354","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.76},{"source":"arxiv:2410.10377","target":"arxiv:2502.21321","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.76},{"source":"arxiv:2405.08295","target":"arxiv:2309.05767","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2405.08295","target":"arxiv:2506.08967","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2405.08295","target":"arxiv:2304.12995","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2405.08295","target":"arxiv:2507.02768","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2306.09265","target":"arxiv:2307.06281","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2306.09265","target":"arxiv:2501.02189","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2306.09265","target":"arxiv:2310.03744","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2306.09265","target":"arxiv:2404.07214","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2002.04501","target":"arxiv:2105.11203","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2002.04501","target":"arxiv:2108.13343","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2002.04501","target":"arxiv:1906.10184","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2002.04501","target":"arxiv:2201.06387","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2309.13638","target":"arxiv:2502.21321","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2309.13638","target":"arxiv:2403.08763","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2309.13638","target":"arxiv:2403.04121","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2309.13638","target":"arxiv:2401.07851","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1903.11027","target":"arxiv:2411.13112","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.74},{"source":"arxiv:1903.11027","target":"arxiv:2506.09042","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.72},{"source":"arxiv:1903.11027","target":"arxiv:2509.06266","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.71},{"source":"arxiv:1903.11027","target":"arxiv:2311.01017","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.71},{"source":"arxiv:2402.12275","target":"arxiv:2405.15383","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2402.12275","target":"arxiv:2606.09032","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2402.12275","target":"arxiv:2511.22904","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2402.12275","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2410.07484","target":"arxiv:2504.15785","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"arxiv:2410.07484","target":"arxiv:2506.09171","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2410.07484","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2410.07484","target":"arxiv:2506.06725","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2001.10371","target":"arxiv:2103.04723","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2001.10371","target":"arxiv:2301.08861","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2001.10371","target":"arxiv:2112.07103","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2001.10371","target":"arxiv:2212.13472","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"1x.tech:d48ab282348b","target":"arxiv:2510.07092","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"1x.tech:d48ab282348b","target":"arxiv:2410.18072","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"1x.tech:d48ab282348b","target":"fi.ee.tsinghua.edu.cn:1c1fbee23e33","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"1x.tech:d48ab282348b","target":"arxiv:2505.09694","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2509.06266","target":"arxiv:2602.03916","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2509.06266","target":"arxiv:2511.15722","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2509.06266","target":"arxiv:2510.08531","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2509.06266","target":"arxiv:2411.13112","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2410.11767","target":"arxiv:2309.08600","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2410.11767","target":"arxiv:2406.11944","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2410.11767","target":"transformer-circuits.pub:e2306116790e","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2410.11767","target":"arxiv:2505.02074","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2602.12520","target":"arxiv:2406.15836","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2602.12520","target":"arxiv:2305.17198","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2602.12520","target":"arxiv:2410.02664","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2602.12520","target":"arxiv:2208.01769","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2606.15874","target":"arxiv:2308.03688","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2606.15874","target":"arxiv:2509.02547","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2606.15874","target":"arxiv:2412.14161","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2606.15874","target":"arxiv:2601.01885","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2511.22904","target":"arxiv:2308.01399","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2511.22904","target":"arxiv:2402.01695","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2511.22904","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2511.22904","target":"arxiv:2305.10626","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"philpapers.org:e7fc6fd114b8","target":"philpapers.org:7b684275d20a","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.94},{"source":"philpapers.org:e7fc6fd114b8","target":"philpapers.org:8c0576aa7f12","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"philpapers.org:e7fc6fd114b8","target":"philpapers.org:9a9c4dcb81d9","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"philpapers.org:e7fc6fd114b8","target":"philpapers.org:4f7ed9918446","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2402.08268","target":"arxiv:2505.21996","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2402.08268","target":"arxiv:2406.09455","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2402.08268","target":"arxiv:2312.14125","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2402.08268","target":"arxiv:2407.12036","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2601.17067","target":"arxiv:2603.28489","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2601.17067","target":"arxiv:2511.08585","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2601.17067","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2601.17067","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2402.06590","target":"nature.com:ad56347f1921","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2402.06590","target":"arxiv:2105.07308","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2402.06590","target":"arxiv:2512.02419","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2402.06590","target":"arxiv:2202.01682","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2211.10831","target":"arxiv:2403.00504","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2211.10831","target":"arxiv:2301.08243","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2211.10831","target":"x.com:351692e95fe5","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2211.10831","target":"arxiv:2404.08471","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2403.14403","target":"arxiv:2404.00610","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2403.14403","target":"arxiv:2405.06211","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2403.14403","target":"arxiv:2407.02485","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2403.14403","target":"arxiv:2505.14146","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2411.04118","target":"arxiv:2512.00597","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2411.04118","target":"arxiv:2502.21321","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2411.04118","target":"arxiv:2404.07214","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2411.04118","target":"arxiv:2507.05201","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2312.15796","target":"nature.com:386c3df42560","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"arxiv:2312.15796","target":"arxiv:2212.12794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2312.15796","target":"arxiv:2306.14066","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2312.15796","target":"arxiv:2405.05925","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2311.17918","target":"arxiv:2502.10498","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2311.17918","target":"arxiv:2403.02622","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2311.17918","target":"arxiv:2309.09777","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2311.17918","target":"arxiv:2405.17398","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1911.10601","target":"arxiv:1904.08149","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:1911.10601","target":"arxiv:1909.10863","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:1911.10601","target":"arxiv:2001.11841","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:1911.10601","target":"arxiv:2505.19867","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2012.04293","target":"arxiv:1910.01442","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2012.04293","target":"arxiv:2510.04978","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2012.04293","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2012.04293","target":"arxiv:2505.02074","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2406.07507","target":"arxiv:2410.02548","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2406.07507","target":"arxiv:2210.02747","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2406.07507","target":"arxiv:2304.14772","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2406.07507","target":"arxiv:2209.15571","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2603.20927","target":"arxiv:2508.12027","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2603.20927","target":"arxiv:2505.19867","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2603.20927","target":"arxiv:1909.10863","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2603.20927","target":"arxiv:1709.02341","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2508.08762","target":"arxiv:2006.04182","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2508.08762","target":"arxiv:2202.09467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2508.08762","target":"arxiv:2406.16062","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2508.08762","target":"arxiv:2103.03725","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"nature.com:c3aa5b0d9273","target":"nature.com:ef956cce5b54","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"nature.com:c3aa5b0d9273","target":"arxiv:2202.01682","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"nature.com:c3aa5b0d9273","target":"nature.com:e173b62796c2","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"nature.com:c3aa5b0d9273","target":"nature.com:ad56347f1921","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2511.15722","target":"arxiv:2510.25760","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.94},{"source":"arxiv:2511.15722","target":"arxiv:2602.03916","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2511.15722","target":"arxiv:2510.08531","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2511.15722","target":"arxiv:2509.06266","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2402.02716","target":"arxiv:2603.07670","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2402.02716","target":"arxiv:2506.09171","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2402.02716","target":"arxiv:2411.06559","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2402.02716","target":"arxiv:2502.12130","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2311.03658","target":"arxiv:2406.01506","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2311.03658","target":"arxiv:2307.15054","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2311.03658","target":"arxiv:2310.02207","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2311.03658","target":"arxiv:2309.00941","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:1910.01442","target":"arxiv:2012.04293","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1910.01442","target":"arxiv:2510.26802","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:1910.01442","target":"arxiv:2502.11831","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:1910.01442","target":"arxiv:2509.12387","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2606.09032","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2606.09032","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2606.09032","target":"arxiv:2308.03688","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2606.09032","target":"arxiv:2411.06559","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2406.06385","target":"arxiv:2305.17888","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2406.06385","target":"arxiv:2303.08302","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2406.06385","target":"arxiv:2504.05352","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2406.06385","target":"arxiv:2308.09723","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2211.10438","target":"arxiv:2312.03788","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.94},{"source":"arxiv:2211.10438","target":"arxiv:2308.09723","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2211.10438","target":"arxiv:2305.17888","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2211.10438","target":"arxiv:2303.08302","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2305.16291","target":"arxiv:2206.08853","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2305.16291","target":"arxiv:2109.06780","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2305.16291","target":"arxiv:2602.15400","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2305.16291","target":"arxiv:2305.10626","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2404.00610","target":"arxiv:2403.14403","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2404.00610","target":"arxiv:2312.10997","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2404.00610","target":"arxiv:2505.14146","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2404.00610","target":"arxiv:2405.06211","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2105.04906","target":"arxiv:1711.00937","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2105.04906","target":"arxiv:2307.12698","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2105.04906","target":"arxiv:2404.08471","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2105.04906","target":"arxiv:2301.08243","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"fi.ee.tsinghua.edu.cn:1c1fbee23e33","target":"arxiv:2510.16732","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"fi.ee.tsinghua.edu.cn:1c1fbee23e33","target":"arxiv:2511.02097","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"fi.ee.tsinghua.edu.cn:1c1fbee23e33","target":"arxiv:2605.12090","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"fi.ee.tsinghua.edu.cn:1c1fbee23e33","target":"arxiv:2507.00917","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2502.11946","target":"arxiv:2506.08967","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2502.11946","target":"arxiv:2304.12995","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2502.11946","target":"arxiv:2405.08295","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2502.11946","target":"arxiv:2309.05767","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"technologyreview.com:6e6b17c22c9b","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"technologyreview.com:6e6b17c22c9b","target":"reply.com:5748177416db","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"technologyreview.com:6e6b17c22c9b","target":"thesequence.substack.com:0f25bda15c4b","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"technologyreview.com:6e6b17c22c9b","target":"arxiv:2411.14499","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2309.09777","target":"arxiv:2403.06845","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2309.09777","target":"arxiv:2502.10498","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2309.09777","target":"arxiv:2403.02622","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2309.09777","target":"arxiv:2311.17918","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2506.22567","target":"arxiv:2303.00915","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2506.22567","target":"arxiv:2210.04133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2506.22567","target":"arxiv:2512.00597","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2506.22567","target":"arxiv:2109.01134","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2411.15594","target":"arxiv:2507.10535","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2411.15594","target":"arxiv:2306.05685","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2411.15594","target":"arxiv:2305.17926","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2411.15594","target":"arxiv:2502.14739","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1805.07526","target":"arxiv:2202.09467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:1805.07526","target":"arxiv:2103.03725","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:1805.07526","target":"arxiv:2207.12316","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:1805.07526","target":"arxiv:2308.07870","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2402.12451","target":"arxiv:2411.06284","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2402.12451","target":"arxiv:2404.07214","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2402.12451","target":"arxiv:2501.02189","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2402.12451","target":"arxiv:2407.00118","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2206.08853","target":"arxiv:2109.06780","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2206.08853","target":"arxiv:2410.02664","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2206.08853","target":"arxiv:2301.04104","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2206.08853","target":"arxiv:2208.01769","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2306.15668","target":"arxiv:2106.08261","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2306.15668","target":"arxiv:2501.09038","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2306.15668","target":"arxiv:2501.16411","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2306.15668","target":"arxiv:2406.03520","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2403.08763","target":"arxiv:2502.21321","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2403.08763","target":"arxiv:2308.12284","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2403.08763","target":"arxiv:2403.08540","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2403.08763","target":"arxiv:2401.02954","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2406.07472","target":"arxiv:2605.22882","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2406.07472","target":"openreview.net:c15700430713","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2406.07472","target":"arxiv:2204.03458","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2406.07472","target":"arxiv:2605.12957","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2410.00161","target":"arxiv:2606.24467","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2410.00161","target":"arxiv:2508.06297","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2410.00161","target":"arxiv:2509.00579","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2410.00161","target":"arxiv:2407.18003","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:0910.3570","target":"arxiv:1011.0063","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:0910.3570","target":"arxiv:1012.5151","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:0910.3570","target":"arxiv:1911.03537","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:0910.3570","target":"arxiv:2208.06726","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2210.13382","target":"arxiv:2309.00941","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2210.13382","target":"arxiv:2403.15498","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2210.13382","target":"arxiv:2305.11169","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2210.13382","target":"arxiv:2605.23972","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2308.01399","target":"arxiv:2511.22904","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2308.01399","target":"arxiv:2305.10626","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2308.01399","target":"arxiv:2402.01695","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2308.01399","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"philpapers.org:f2d0d9f7c8fe","target":"philpapers.org:8c0576aa7f12","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"philpapers.org:f2d0d9f7c8fe","target":"philpapers.org:7b684275d20a","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"philpapers.org:f2d0d9f7c8fe","target":"philpapers.org:e7fc6fd114b8","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"philpapers.org:f2d0d9f7c8fe","target":"philarchive.org:04bcc5cbd425","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2509.20021","target":"arxiv:2407.06886","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2509.20021","target":"arxiv:2506.22355","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2509.20021","target":"arxiv:2507.00917","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2509.20021","target":"arxiv:2510.16732","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:1711.00937","target":"arxiv:2404.08471","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:1711.00937","target":"arxiv:2308.07870","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:1711.00937","target":"arxiv:1312.6114","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:1711.00937","target":"arxiv:2006.04182","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2410.08893","target":"arxiv:2107.08241","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2410.08893","target":"arxiv:2310.09615","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2410.08893","target":"arxiv:2303.07109","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2410.08893","target":"arxiv:2406.15836","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:cs/0405001","target":"arxiv:2401.03804","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.66},{"source":"arxiv:cs/0405001","target":"arxiv:2510.21219","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.63},{"source":"arxiv:cs/0405001","target":"arxiv:2604.22748","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.63},{"source":"arxiv:cs/0405001","target":"arxiv:2605.16689","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.63},{"source":"pi.website:ca5c1a691358","target":"arxiv:2606.06556","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"pi.website:ca5c1a691358","target":"arxiv:2508.13073","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"pi.website:ca5c1a691358","target":"arxiv:2603.18532","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"pi.website:ca5c1a691358","target":"arxiv:2307.15818","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1909.10863","target":"arxiv:2508.12027","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:1909.10863","target":"arxiv:2001.11841","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:1909.10863","target":"arxiv:1911.10601","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:1909.10863","target":"arxiv:1904.08149","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2304.07193","target":"arxiv:2404.08471","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2304.07193","target":"arxiv:2406.06973","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2304.07193","target":"arxiv:2109.01134","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2304.07193","target":"arxiv:2103.00020","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2510.17111","target":"arxiv:2508.13073","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2510.17111","target":"arxiv:2502.05855","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2510.17111","target":"arxiv:2507.12440","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2510.17111","target":"arxiv:2603.18532","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2603.21354","target":"arxiv:2510.03847","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2603.21354","target":"arxiv:2603.20224","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2603.21354","target":"arxiv:2402.02716","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2603.21354","target":"arxiv:2602.24044","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2511.02097","target":"fi.ee.tsinghua.edu.cn:1c1fbee23e33","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2511.02097","target":"arxiv:2403.02622","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2511.02097","target":"arxiv:2605.12090","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2511.02097","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2510.25760","target":"arxiv:2511.15722","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.94},{"source":"arxiv:2510.25760","target":"arxiv:2602.03916","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2510.25760","target":"arxiv:2509.06266","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2510.25760","target":"arxiv:2510.08531","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2505.14146","target":"arxiv:2404.00610","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2505.14146","target":"arxiv:2405.06211","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2505.14146","target":"arxiv:2005.11401","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2505.14146","target":"arxiv:2407.02485","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2303.07109","target":"arxiv:2209.00588","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2303.07109","target":"arxiv:2310.09615","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2303.07109","target":"arxiv:2406.15836","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2303.07109","target":"arxiv:1803.10122","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2506.01622","target":"arxiv:2511.22904","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2506.01622","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2506.01622","target":"arxiv:2305.10626","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2506.01622","target":"arxiv:1803.10122","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2303.03378","target":"arxiv:2510.17111","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2303.03378","target":"arxiv:2508.13073","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2303.03378","target":"arxiv:2507.12440","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2303.03378","target":"arxiv:2402.12451","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2406.15836","target":"arxiv:2602.12520","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2406.15836","target":"arxiv:2303.07109","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2406.15836","target":"arxiv:2410.02664","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2406.15836","target":"arxiv:2209.00588","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2305.14909","target":"arxiv:2305.14992","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2305.14909","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2305.14909","target":"arxiv:2403.04121","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2305.14909","target":"arxiv:2502.13092","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2406.06485","target":"arxiv:2411.08794","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2406.06485","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2406.06485","target":"arxiv:2605.23972","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2406.06485","target":"arxiv:2606.09032","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2504.15785","target":"arxiv:2410.07484","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"arxiv:2504.15785","target":"arxiv:2411.04983","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2504.15785","target":"arxiv:2506.06725","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2504.15785","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2309.07870","target":"arxiv:2402.16823","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2309.07870","target":"arxiv:2308.03688","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2309.07870","target":"arxiv:2502.12130","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2309.07870","target":"arxiv:2503.16416","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2201.11903","target":"arxiv:2205.11916","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2201.11903","target":"arxiv:2502.03373","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2201.11903","target":"arxiv:2305.10601","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2201.11903","target":"arxiv:2502.18080","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2603.03251","target":"arxiv:2411.13157","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2603.03251","target":"arxiv:2511.20340","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2603.03251","target":"arxiv:2401.07851","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2603.03251","target":"arxiv:2402.01528","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2507.05169","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2507.05169","target":"arxiv:2601.15533","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2507.05169","target":"arxiv:2411.14499","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2507.05169","target":"arxiv:2510.21219","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2307.06281","target":"arxiv:2306.09265","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2307.06281","target":"arxiv:2501.02189","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2307.06281","target":"arxiv:2307.16125","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2307.06281","target":"arxiv:2402.12451","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2412.14161","target":"neurips.cc:9c7e4f09bb35","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.96},{"source":"arxiv:2412.14161","target":"arxiv:2308.03688","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2412.14161","target":"arxiv:2503.16416","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2412.14161","target":"arxiv:2401.13178","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2307.12698","target":"arxiv:2301.08243","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2307.12698","target":"arxiv:2404.08471","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2307.12698","target":"arxiv:2506.09985","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2307.12698","target":"arxiv:2211.10831","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:1811.04551","target":"arxiv:2206.04114","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:1811.04551","target":"arxiv:2502.14819","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:1811.04551","target":"arxiv:2107.08241","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:1811.04551","target":"arxiv:1803.10122","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2103.00020","target":"arxiv:2109.01134","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2103.00020","target":"arxiv:2406.06973","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2103.00020","target":"arxiv:2110.04544","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2103.00020","target":"arxiv:2412.07077","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2506.18701","target":"arxiv:2412.00887","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2506.18701","target":"arxiv:2412.03568","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2506.18701","target":"arxiv:2505.21996","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2506.18701","target":"arxiv:2411.00769","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2503.15168","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2503.15168","target":"arxiv:2512.02419","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2503.15168","target":"technologyreview.com:6e6b17c22c9b","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2503.15168","target":"arxiv:2511.12239","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2510.19788","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2510.19788","target":"arxiv:2506.01622","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2510.19788","target":"arxiv:2503.15168","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2510.19788","target":"technologyreview.com:6e6b17c22c9b","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"nature.com:ef956cce5b54","target":"nature.com:c3aa5b0d9273","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"nature.com:ef956cce5b54","target":"nature.com:e173b62796c2","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"nature.com:ef956cce5b54","target":"nature.com:ad56347f1921","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"nature.com:ef956cce5b54","target":"arxiv:1805.09042","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2306.10754","target":"arxiv:2202.03771","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2306.10754","target":"arxiv:2203.03417","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2306.10754","target":"arxiv:2112.07103","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2306.10754","target":"arxiv:2103.04723","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2102.11107","target":"arxiv:2509.12387","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2102.11107","target":"arxiv:2505.02074","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2102.11107","target":"arxiv:2506.23068","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2102.11107","target":"arxiv:2410.11767","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2411.13157","target":"arxiv:2401.07851","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"arxiv:2411.13157","target":"arxiv:2511.20340","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2411.13157","target":"arxiv:2603.03251","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2411.13157","target":"arxiv:2402.01528","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2001.11841","target":"arxiv:1709.02341","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2001.11841","target":"arxiv:2508.12027","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2001.11841","target":"arxiv:1904.08149","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2001.11841","target":"arxiv:1909.10863","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2412.17254","target":"arxiv:2305.18264","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2412.17254","target":"arxiv:2204.03458","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2412.17254","target":"arxiv:2505.21996","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2412.17254","target":"arxiv:2510.04999","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2606.10749","target":"arxiv:2503.16416","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2606.10749","target":"arxiv:2603.07670","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2606.10749","target":"arxiv:2402.10196","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2606.10749","target":"arxiv:2308.03688","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2407.02485","target":"arxiv:2404.00610","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2407.02485","target":"arxiv:2403.14403","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2407.02485","target":"arxiv:2312.10997","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2407.02485","target":"arxiv:2405.06211","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2503.02854","target":"arxiv:2305.11169","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2503.02854","target":"arxiv:2402.10644","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2503.02854","target":"arxiv:2210.13382","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2503.02854","target":"arxiv:2309.00941","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2406.03520","target":"arxiv:2503.06800","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"arxiv:2406.03520","target":"arxiv:2501.09038","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2406.03520","target":"arxiv:2411.02385","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2406.03520","target":"arxiv:2506.09849","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2604.22748","target":"arxiv:2606.00133","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2604.22748","target":"arxiv:2511.02097","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2604.22748","target":"technologyreview.com:6e6b17c22c9b","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2604.22748","target":"arxiv:2606.09032","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2101.03961","target":"arxiv:2202.08906","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2101.03961","target":"arxiv:2403.08763","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2101.03961","target":"arxiv:2308.12284","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2101.03961","target":"arxiv:2406.11944","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2310.09615","target":"arxiv:2209.00588","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2310.09615","target":"arxiv:2303.07109","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2310.09615","target":"arxiv:2406.15836","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2310.09615","target":"arxiv:1903.00374","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2402.10644","target":"arxiv:2405.10480","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2402.10644","target":"arxiv:1706.03762","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2402.10644","target":"arxiv:2503.02854","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2402.10644","target":"arxiv:2303.07109","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:2306.02572","target":"openreview.net:3f48a33ff9bf","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2306.02572","target":"arxiv:2505.19867","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2306.02572","target":"arxiv:2503.15168","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2306.02572","target":"arxiv:2407.00118","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:2406.01506","target":"arxiv:2311.03658","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2406.01506","target":"arxiv:2307.15054","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2406.01506","target":"arxiv:2310.02207","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.78},{"source":"arxiv:2406.01506","target":"arxiv:2309.00941","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.76},{"source":"arxiv:2312.10997","target":"arxiv:2405.06211","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.95},{"source":"arxiv:2312.10997","target":"arxiv:2401.15884","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2312.10997","target":"arxiv:2005.11401","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2312.10997","target":"arxiv:2404.00610","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:1911.03537","target":"arxiv:1011.0063","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:1911.03537","target":"arxiv:1012.5151","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:1911.03537","target":"arxiv:0910.3570","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.8},{"source":"arxiv:1911.03537","target":"arxiv:2208.06726","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.76},{"source":"arxiv:2301.08861","target":"arxiv:2103.04723","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2301.08861","target":"arxiv:2112.07103","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2301.08861","target":"arxiv:2001.10371","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2301.08861","target":"arxiv:2212.13472","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2111.06377","target":"arxiv:2410.11767","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2111.06377","target":"arxiv:2301.08243","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2111.06377","target":"arxiv:2305.07017","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2111.06377","target":"arxiv:2304.07193","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.77},{"source":"arxiv:2311.17593","target":"arxiv:2511.22904","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2311.17593","target":"arxiv:2305.10626","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2311.17593","target":"arxiv:2402.01695","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2311.17593","target":"arxiv:2510.19818","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2601.01885","target":"arxiv:2603.07670","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2601.01885","target":"arxiv:2509.02547","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2601.01885","target":"arxiv:2506.09171","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2601.01885","target":"arxiv:2402.02716","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2005.11401","target":"arxiv:2312.10997","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2005.11401","target":"arxiv:2405.06211","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2005.11401","target":"arxiv:2404.00610","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2005.11401","target":"arxiv:2505.14146","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2305.17888","target":"arxiv:2406.06385","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.92},{"source":"arxiv:2305.17888","target":"arxiv:2211.10438","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2305.17888","target":"arxiv:2303.08302","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2305.17888","target":"arxiv:2504.05352","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2510.08531","target":"arxiv:2511.15722","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2510.08531","target":"arxiv:2602.03916","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2510.08531","target":"arxiv:2509.06266","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2510.08531","target":"arxiv:2510.25760","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:1803.07616","target":"arxiv:2506.09849","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:1803.07616","target":"arxiv:2502.11831","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:1803.07616","target":"arxiv:2106.08261","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.81},{"source":"arxiv:1803.07616","target":"arxiv:2501.16411","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.79},{"source":"arxiv:2305.16653","target":"arxiv:2302.01560","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2305.16653","target":"arxiv:2305.14992","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2305.16653","target":"arxiv:2402.02716","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2305.16653","target":"arxiv:2305.14909","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2507.02768","target":"arxiv:2506.08967","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2507.02768","target":"arxiv:2309.05767","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2507.02768","target":"arxiv:2405.08295","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.83},{"source":"arxiv:2507.02768","target":"arxiv:2304.12995","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2605.12957","target":"huggingface.co:9f24e732f5fe","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2605.12957","target":"arxiv:2605.22882","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2605.12957","target":"arxiv:2603.13910","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2605.12957","target":"arxiv:2204.03458","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2305.11169","target":"arxiv:2210.13382","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2305.11169","target":"arxiv:2106.00737","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2305.11169","target":"arxiv:2511.22904","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2305.11169","target":"the-decoder.com:97885f462d1f","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.82},{"source":"arxiv:2409.10644","target":"arxiv:2510.12966","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.93},{"source":"arxiv:2409.10644","target":"arxiv:2402.01528","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.91},{"source":"arxiv:2409.10644","target":"arxiv:2411.13157","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.9},{"source":"arxiv:2409.10644","target":"arxiv:2401.07851","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2311.17042","target":"arxiv:2405.05967","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.89},{"source":"arxiv:2311.17042","target":"arxiv:2105.05233","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.88},{"source":"arxiv:2311.17042","target":"arxiv:2406.02347","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.87},{"source":"arxiv:2311.17042","target":"arxiv:2112.10752","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2508.16745","target":"arxiv:2502.18080","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.86},{"source":"arxiv:2508.16745","target":"arxiv:2502.03373","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.85},{"source":"arxiv:2508.16745","target":"arxiv:2503.16419","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"arxiv:2508.16745","target":"arxiv:2504.10903","relation":"semantic","confidence":"EXTRACTED","confidence_score":0.84},{"source":"0a7193294bbb5d1f9097552f2242f34f","target":"concept_t30","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"0a7193294bbb5d1f9097552f2242f34f","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1706.03762","target":"0a7193294bbb5d1f9097552f2242f34f","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1709.06560","target":"0a7193294bbb5d1f9097552f2242f34f","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1711.00832","target":"0a7193294bbb5d1f9097552f2242f34f","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1803.07616","target":"0a7193294bbb5d1f9097552f2242f34f","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"c5c0a185adae51f49af390e3d5f679c6","target":"concept_t0","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1705.09156","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1706.02275","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1706.03741","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1709.02341","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"00ed333d471b517cacd77ef474592c37","target":"concept_t17","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1510.00056","target":"00ed333d471b517cacd77ef474592c37","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1706.02275","target":"00ed333d471b517cacd77ef474592c37","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1706.03741","target":"00ed333d471b517cacd77ef474592c37","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1708.00133","target":"00ed333d471b517cacd77ef474592c37","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1709.02341","target":"00ed333d471b517cacd77ef474592c37","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"1ea0320008f054d6a48355261be0862a","target":"concept_t0","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"1ea0320008f054d6a48355261be0862a","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1705.09156","target":"1ea0320008f054d6a48355261be0862a","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1706.02275","target":"1ea0320008f054d6a48355261be0862a","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1706.03741","target":"1ea0320008f054d6a48355261be0862a","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1708.00133","target":"1ea0320008f054d6a48355261be0862a","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"e2e9917abed652b59f2892d75ccdec63","target":"concept_t19","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"aclanthology.org:8f7acb76ba1d","target":"e2e9917abed652b59f2892d75ccdec63","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1706.02275","target":"e2e9917abed652b59f2892d75ccdec63","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1706.03762","target":"e2e9917abed652b59f2892d75ccdec63","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1708.00133","target":"e2e9917abed652b59f2892d75ccdec63","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1910.08210","target":"e2e9917abed652b59f2892d75ccdec63","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"ef9d9f01afc951daaba0da0ceaeb4988","target":"concept_t30","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1711.00937","target":"ef9d9f01afc951daaba0da0ceaeb4988","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1803.07616","target":"ef9d9f01afc951daaba0da0ceaeb4988","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1903.11027","target":"ef9d9f01afc951daaba0da0ceaeb4988","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1908.05656","target":"ef9d9f01afc951daaba0da0ceaeb4988","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1910.01442","target":"ef9d9f01afc951daaba0da0ceaeb4988","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"5e88711b4de2546296d173f144a8ca38","target":"concept_t17","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"5e88711b4de2546296d173f144a8ca38","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1709.02341","target":"5e88711b4de2546296d173f144a8ca38","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1711.00937","target":"5e88711b4de2546296d173f144a8ca38","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1712.04602","target":"5e88711b4de2546296d173f144a8ca38","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1801.04819","target":"5e88711b4de2546296d173f144a8ca38","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"6aeb0ae745b350e18820468d8ce4df3e","target":"concept_t26","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"6aeb0ae745b350e18820468d8ce4df3e","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1510.00056","target":"6aeb0ae745b350e18820468d8ce4df3e","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1706.02275","target":"6aeb0ae745b350e18820468d8ce4df3e","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1706.03741","target":"6aeb0ae745b350e18820468d8ce4df3e","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1708.00133","target":"6aeb0ae745b350e18820468d8ce4df3e","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"1fbb24e7e94e58e18899bd2b06c8fed4","target":"concept_t0","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"1fbb24e7e94e58e18899bd2b06c8fed4","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1011.0063","target":"1fbb24e7e94e58e18899bd2b06c8fed4","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1701.06970","target":"1fbb24e7e94e58e18899bd2b06c8fed4","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1709.02341","target":"1fbb24e7e94e58e18899bd2b06c8fed4","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1711.00937","target":"1fbb24e7e94e58e18899bd2b06c8fed4","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"b76fb931c1c75ba7964ec119fc2a0b34","target":"concept_t30","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2005.11401","target":"b76fb931c1c75ba7964ec119fc2a0b34","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2101.03961","target":"b76fb931c1c75ba7964ec119fc2a0b34","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2102.01192","target":"b76fb931c1c75ba7964ec119fc2a0b34","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2103.00020","target":"b76fb931c1c75ba7964ec119fc2a0b34","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2106.00737","target":"b76fb931c1c75ba7964ec119fc2a0b34","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"578297fe62465056b5d74185e59b4615","target":"concept_t0","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"578297fe62465056b5d74185e59b4615","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1706.03741","target":"578297fe62465056b5d74185e59b4615","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1709.06560","target":"578297fe62465056b5d74185e59b4615","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1801.04819","target":"578297fe62465056b5d74185e59b4615","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1811.04551","target":"578297fe62465056b5d74185e59b4615","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"b091ad18c2395679bfef4cda988e8378","target":"concept_t32","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1711.00937","target":"b091ad18c2395679bfef4cda988e8378","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1907.05600","target":"b091ad18c2395679bfef4cda988e8378","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2006.11239","target":"b091ad18c2395679bfef4cda988e8378","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2105.05233","target":"b091ad18c2395679bfef4cda988e8378","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2112.07103","target":"b091ad18c2395679bfef4cda988e8378","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"3309b4841a9456de8a44882428576753","target":"concept_t30","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1910.01442","target":"3309b4841a9456de8a44882428576753","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2010.11929","target":"3309b4841a9456de8a44882428576753","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2012.04293","target":"3309b4841a9456de8a44882428576753","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2103.00020","target":"3309b4841a9456de8a44882428576753","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2109.01134","target":"3309b4841a9456de8a44882428576753","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"3e5f643d839e5aa2b8723a46fba6a2a6","target":"concept_t29","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"3e5f643d839e5aa2b8723a46fba6a2a6","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1711.00937","target":"3e5f643d839e5aa2b8723a46fba6a2a6","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1803.07616","target":"3e5f643d839e5aa2b8723a46fba6a2a6","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1803.10122","target":"3e5f643d839e5aa2b8723a46fba6a2a6","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1811.04551","target":"3e5f643d839e5aa2b8723a46fba6a2a6","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"cf4ebe8a6fd0551fa06e22cccb6525d9","target":"concept_t28","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1712.04602","target":"cf4ebe8a6fd0551fa06e22cccb6525d9","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1803.10122","target":"cf4ebe8a6fd0551fa06e22cccb6525d9","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1805.09042","target":"cf4ebe8a6fd0551fa06e22cccb6525d9","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1805.09176","target":"cf4ebe8a6fd0551fa06e22cccb6525d9","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1807.11819","target":"cf4ebe8a6fd0551fa06e22cccb6525d9","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"afbadd921ac75a03a7cb4ed228efc708","target":"concept_t24","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2101.03961","target":"afbadd921ac75a03a7cb4ed228efc708","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2211.10438","target":"afbadd921ac75a03a7cb4ed228efc708","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2212.13472","target":"afbadd921ac75a03a7cb4ed228efc708","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2303.01469","target":"afbadd921ac75a03a7cb4ed228efc708","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2303.08302","target":"afbadd921ac75a03a7cb4ed228efc708","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"5d706c4d8e3b5c51bf1534e0ac3649be","target":"concept_t16","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1312.6114","target":"5d706c4d8e3b5c51bf1534e0ac3649be","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1805.07526","target":"5d706c4d8e3b5c51bf1534e0ac3649be","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1805.09042","target":"5d706c4d8e3b5c51bf1534e0ac3649be","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1805.09176","target":"5d706c4d8e3b5c51bf1534e0ac3649be","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1807.11819","target":"5d706c4d8e3b5c51bf1534e0ac3649be","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"d613731276125291ae874322d754e3e0","target":"concept_t29","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"d613731276125291ae874322d754e3e0","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1711.00937","target":"d613731276125291ae874322d754e3e0","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2010.02193","target":"d613731276125291ae874322d754e3e0","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2102.11107","target":"d613731276125291ae874322d754e3e0","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2106.08261","target":"d613731276125291ae874322d754e3e0","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"f7d9b5d7ab0e51f8a085ffe40543f902","target":"concept_t15","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1903.11027","target":"f7d9b5d7ab0e51f8a085ffe40543f902","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1907.05600","target":"f7d9b5d7ab0e51f8a085ffe40543f902","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2010.11929","target":"f7d9b5d7ab0e51f8a085ffe40543f902","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2106.08261","target":"f7d9b5d7ab0e51f8a085ffe40543f902","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2111.06377","target":"f7d9b5d7ab0e51f8a085ffe40543f902","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"d5699c11ff3e5f2ea251177377b18fff","target":"concept_t0","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"d5699c11ff3e5f2ea251177377b18fff","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1910.08210","target":"d5699c11ff3e5f2ea251177377b18fff","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1911.11641","target":"d5699c11ff3e5f2ea251177377b18fff","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2204.01691","target":"d5699c11ff3e5f2ea251177377b18fff","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2205.06175","target":"d5699c11ff3e5f2ea251177377b18fff","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"e0312e04abde5e53a1f0b3a14d1422ba","target":"concept_t8","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1706.03741","target":"e0312e04abde5e53a1f0b3a14d1422ba","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2005.14165","target":"e0312e04abde5e53a1f0b3a14d1422ba","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2103.00020","target":"e0312e04abde5e53a1f0b3a14d1422ba","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2208.01769","target":"e0312e04abde5e53a1f0b3a14d1422ba","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2303.00915","target":"e0312e04abde5e53a1f0b3a14d1422ba","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"0bf069239279577ebdb04c2e7f09c1c9","target":"concept_t17","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1803.07616","target":"0bf069239279577ebdb04c2e7f09c1c9","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1803.10122","target":"0bf069239279577ebdb04c2e7f09c1c9","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1812.00922","target":"0bf069239279577ebdb04c2e7f09c1c9","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1903.00374","target":"0bf069239279577ebdb04c2e7f09c1c9","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1903.11027","target":"0bf069239279577ebdb04c2e7f09c1c9","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"5a9f7bc5b9f45daca6a11a434b3f61d9","target":"concept_t28","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1705.09156","target":"5a9f7bc5b9f45daca6a11a434b3f61d9","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1709.02341","target":"5a9f7bc5b9f45daca6a11a434b3f61d9","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1712.04602","target":"5a9f7bc5b9f45daca6a11a434b3f61d9","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1801.04819","target":"5a9f7bc5b9f45daca6a11a434b3f61d9","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1803.07616","target":"5a9f7bc5b9f45daca6a11a434b3f61d9","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"2fe319dfa7aa5c34b3373b94d57bfacc","target":"concept_t26","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1706.02275","target":"2fe319dfa7aa5c34b3373b94d57bfacc","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1706.03741","target":"2fe319dfa7aa5c34b3373b94d57bfacc","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1708.00133","target":"2fe319dfa7aa5c34b3373b94d57bfacc","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1711.00832","target":"2fe319dfa7aa5c34b3373b94d57bfacc","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1711.00937","target":"2fe319dfa7aa5c34b3373b94d57bfacc","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"179eca3284ca53d1909a4865bb5c5f67","target":"concept_t0","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2204.01691","target":"179eca3284ca53d1909a4865bb5c5f67","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2205.06175","target":"179eca3284ca53d1909a4865bb5c5f67","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2212.06817","target":"179eca3284ca53d1909a4865bb5c5f67","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2303.03378","target":"179eca3284ca53d1909a4865bb5c5f67","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2307.15818","target":"179eca3284ca53d1909a4865bb5c5f67","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"41a9b2d7e67f56009ef5f9b25218ae23","target":"concept_t5","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"aclanthology.org:8f7acb76ba1d","target":"41a9b2d7e67f56009ef5f9b25218ae23","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1705.09156","target":"41a9b2d7e67f56009ef5f9b25218ae23","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1801.04819","target":"41a9b2d7e67f56009ef5f9b25218ae23","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1805.09176","target":"41a9b2d7e67f56009ef5f9b25218ae23","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1807.11819","target":"41a9b2d7e67f56009ef5f9b25218ae23","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"55dc64e751fc5cada055cbef1ba6f490","target":"concept_t35","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2305.10601","target":"55dc64e751fc5cada055cbef1ba6f490","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2308.03688","target":"55dc64e751fc5cada055cbef1ba6f490","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2308.16369","target":"55dc64e751fc5cada055cbef1ba6f490","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2309.07870","target":"55dc64e751fc5cada055cbef1ba6f490","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2310.08560","target":"55dc64e751fc5cada055cbef1ba6f490","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"156bd9a2ef6c50b78065af31965a80e8","target":"concept_t17","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1812.00922","target":"156bd9a2ef6c50b78065af31965a80e8","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1903.11027","target":"156bd9a2ef6c50b78065af31965a80e8","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2102.11107","target":"156bd9a2ef6c50b78065af31965a80e8","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2112.10752","target":"156bd9a2ef6c50b78065af31965a80e8","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2208.01769","target":"156bd9a2ef6c50b78065af31965a80e8","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"20e90bf1db27548698a1a21ebe2a12da","target":"concept_t28","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1705.09156","target":"20e90bf1db27548698a1a21ebe2a12da","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1712.04602","target":"20e90bf1db27548698a1a21ebe2a12da","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1801.04819","target":"20e90bf1db27548698a1a21ebe2a12da","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1805.09042","target":"20e90bf1db27548698a1a21ebe2a12da","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1807.11819","target":"20e90bf1db27548698a1a21ebe2a12da","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"dfdc4dcc97d55b138155289efc1f8f15","target":"concept_t32","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1312.6114","target":"dfdc4dcc97d55b138155289efc1f8f15","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1907.05600","target":"dfdc4dcc97d55b138155289efc1f8f15","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2006.11239","target":"dfdc4dcc97d55b138155289efc1f8f15","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2105.05233","target":"dfdc4dcc97d55b138155289efc1f8f15","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2112.10752","target":"dfdc4dcc97d55b138155289efc1f8f15","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"6efe5430e683589693361dbeaf1aa256","target":"concept_t16","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"aclanthology.org:8f7acb76ba1d","target":"6efe5430e683589693361dbeaf1aa256","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1706.03762","target":"6efe5430e683589693361dbeaf1aa256","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1810.04805","target":"6efe5430e683589693361dbeaf1aa256","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1910.10683","target":"6efe5430e683589693361dbeaf1aa256","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1911.11641","target":"6efe5430e683589693361dbeaf1aa256","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"3abafab4d0695bef9da5342372fe06ce","target":"concept_t1","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2206.10498","target":"3abafab4d0695bef9da5342372fe06ce","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2303.11366","target":"3abafab4d0695bef9da5342372fe06ce","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2305.11169","target":"3abafab4d0695bef9da5342372fe06ce","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2312.03788","target":"3abafab4d0695bef9da5342372fe06ce","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2401.02954","target":"3abafab4d0695bef9da5342372fe06ce","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"da6911ef37435c6abb169e201aa5f45d","target":"concept_t26","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1706.03741","target":"da6911ef37435c6abb169e201aa5f45d","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1708.00133","target":"da6911ef37435c6abb169e201aa5f45d","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1709.06560","target":"da6911ef37435c6abb169e201aa5f45d","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1711.00832","target":"da6911ef37435c6abb169e201aa5f45d","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1711.00937","target":"da6911ef37435c6abb169e201aa5f45d","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"6f9da474895254f89a0b8c2c1d528480","target":"concept_t13","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1805.09176","target":"6f9da474895254f89a0b8c2c1d528480","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2102.11107","target":"6f9da474895254f89a0b8c2c1d528480","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2306.02572","target":"6f9da474895254f89a0b8c2c1d528480","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2312.10997","target":"6f9da474895254f89a0b8c2c1d528480","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2402.12451","target":"6f9da474895254f89a0b8c2c1d528480","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"1320c285fbd7589ea3b5acd47bdc24e5","target":"concept_t31","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1810.04805","target":"1320c285fbd7589ea3b5acd47bdc24e5","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1910.10683","target":"1320c285fbd7589ea3b5acd47bdc24e5","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2005.14165","target":"1320c285fbd7589ea3b5acd47bdc24e5","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2101.03961","target":"1320c285fbd7589ea3b5acd47bdc24e5","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2202.08906","target":"1320c285fbd7589ea3b5acd47bdc24e5","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"a4a21ac4385e5f19aae2c6647ebddd70","target":"concept_t26","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:0910.3570","target":"a4a21ac4385e5f19aae2c6647ebddd70","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1012.5151","target":"a4a21ac4385e5f19aae2c6647ebddd70","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1510.00056","target":"a4a21ac4385e5f19aae2c6647ebddd70","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1701.06970","target":"a4a21ac4385e5f19aae2c6647ebddd70","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1712.04602","target":"a4a21ac4385e5f19aae2c6647ebddd70","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"de68e0ca85d358759a695e4fe668563e","target":"concept_t30","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2303.03378","target":"de68e0ca85d358759a695e4fe668563e","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2306.09265","target":"de68e0ca85d358759a695e4fe668563e","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2309.02591","target":"de68e0ca85d358759a695e4fe668563e","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2310.03744","target":"de68e0ca85d358759a695e4fe668563e","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2501.02189","target":"de68e0ca85d358759a695e4fe668563e","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"07c2931e095759c7a2114e8d9c086a6b","target":"concept_t11","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1711.00937","target":"07c2931e095759c7a2114e8d9c086a6b","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2102.01192","target":"07c2931e095759c7a2114e8d9c086a6b","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2304.12995","target":"07c2931e095759c7a2114e8d9c086a6b","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2306.02572","target":"07c2931e095759c7a2114e8d9c086a6b","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2310.08803","target":"07c2931e095759c7a2114e8d9c086a6b","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"3a73876784785037a96bc25790a0f5f4","target":"concept_t29","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1803.07616","target":"3a73876784785037a96bc25790a0f5f4","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2206.08853","target":"3a73876784785037a96bc25790a0f5f4","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2304.07193","target":"3a73876784785037a96bc25790a0f5f4","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2307.12698","target":"3a73876784785037a96bc25790a0f5f4","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2311.05348","target":"3a73876784785037a96bc25790a0f5f4","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"a7af1de80dc55255a1b68d98e0e459d0","target":"concept_t15","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2311.01017","target":"a7af1de80dc55255a1b68d98e0e459d0","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2311.16038","target":"a7af1de80dc55255a1b68d98e0e459d0","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2311.17918","target":"a7af1de80dc55255a1b68d98e0e459d0","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2312.15796","target":"a7af1de80dc55255a1b68d98e0e459d0","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2401.12945","target":"a7af1de80dc55255a1b68d98e0e459d0","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"e4011f27b2e150bea808160e79f649c6","target":"concept_t23","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1805.07526","target":"e4011f27b2e150bea808160e79f649c6","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2010.11929","target":"e4011f27b2e150bea808160e79f649c6","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2103.00020","target":"e4011f27b2e150bea808160e79f649c6","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2105.04906","target":"e4011f27b2e150bea808160e79f649c6","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2109.01134","target":"e4011f27b2e150bea808160e79f649c6","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"25f5883a505b5c90b3288786468e427b","target":"concept_t27","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2305.15074","target":"25f5883a505b5c90b3288786468e427b","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2305.18654","target":"25f5883a505b5c90b3288786468e427b","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2401.02954","target":"25f5883a505b5c90b3288786468e427b","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2410.05229","target":"25f5883a505b5c90b3288786468e427b","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2502.12130","target":"25f5883a505b5c90b3288786468e427b","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"11e409aa7db05ac6926994cc5efd7a4e","target":"concept_t23","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1805.07526","target":"11e409aa7db05ac6926994cc5efd7a4e","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1807.11819","target":"11e409aa7db05ac6926994cc5efd7a4e","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2006.04182","target":"11e409aa7db05ac6926994cc5efd7a4e","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2010.11929","target":"11e409aa7db05ac6926994cc5efd7a4e","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2102.11107","target":"11e409aa7db05ac6926994cc5efd7a4e","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"8b61b4feb56658e4877bb944c345a611","target":"concept_t0","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2006.15134","target":"8b61b4feb56658e4877bb944c345a611","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2212.06817","target":"8b61b4feb56658e4877bb944c345a611","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2411.04983","target":"8b61b4feb56658e4877bb944c345a611","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2507.00917","target":"8b61b4feb56658e4877bb944c345a611","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2510.19818","target":"8b61b4feb56658e4877bb944c345a611","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"5a237ad74e26549290d5b57774b015da","target":"concept_t24","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2211.10438","target":"5a237ad74e26549290d5b57774b015da","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2308.16369","target":"5a237ad74e26549290d5b57774b015da","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2312.03788","target":"5a237ad74e26549290d5b57774b015da","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2412.18106","target":"5a237ad74e26549290d5b57774b015da","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2503.20552","target":"5a237ad74e26549290d5b57774b015da","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"d8c361b6739f56a582950c29041b3153","target":"concept_t13","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2102.08363","target":"d8c361b6739f56a582950c29041b3153","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2303.00915","target":"d8c361b6739f56a582950c29041b3153","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2407.06886","target":"d8c361b6739f56a582950c29041b3153","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2411.04118","target":"d8c361b6739f56a582950c29041b3153","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2411.06284","target":"d8c361b6739f56a582950c29041b3153","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"29e518447cab5ffeab46a89a0e892ece","target":"concept_t9","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2305.17888","target":"29e518447cab5ffeab46a89a0e892ece","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2401.07851","target":"29e518447cab5ffeab46a89a0e892ece","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2402.01528","target":"29e518447cab5ffeab46a89a0e892ece","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2405.10480","target":"29e518447cab5ffeab46a89a0e892ece","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2409.10644","target":"29e518447cab5ffeab46a89a0e892ece","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"6cb325eebf8b5cf1b32fd77c46351199","target":"concept_t35","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2305.05383","target":"6cb325eebf8b5cf1b32fd77c46351199","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2402.09664","target":"6cb325eebf8b5cf1b32fd77c46351199","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2403.16971","target":"6cb325eebf8b5cf1b32fd77c46351199","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2405.06211","target":"6cb325eebf8b5cf1b32fd77c46351199","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2411.15594","target":"6cb325eebf8b5cf1b32fd77c46351199","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"33ec7dab157f55ae97bffe3facf6a915","target":"concept_t33","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2305.15074","target":"33ec7dab157f55ae97bffe3facf6a915","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2307.13692","target":"33ec7dab157f55ae97bffe3facf6a915","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2406.06485","target":"33ec7dab157f55ae97bffe3facf6a915","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2408.03943","target":"33ec7dab157f55ae97bffe3facf6a915","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2410.05229","target":"33ec7dab157f55ae97bffe3facf6a915","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"acab376f7b675f42a7378c3c877df082","target":"concept_t16","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2006.04182","target":"acab376f7b675f42a7378c3c877df082","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2106.00737","target":"acab376f7b675f42a7378c3c877df082","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2202.08906","target":"acab376f7b675f42a7378c3c877df082","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2307.15054","target":"acab376f7b675f42a7378c3c877df082","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2402.10644","target":"acab376f7b675f42a7378c3c877df082","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"ad7614ff9d335520b4e20c7315f706c8","target":"concept_t15","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2210.07729","target":"ad7614ff9d335520b4e20c7315f706c8","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2305.10626","target":"ad7614ff9d335520b4e20c7315f706c8","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2310.01361","target":"ad7614ff9d335520b4e20c7315f706c8","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2406.06485","target":"ad7614ff9d335520b4e20c7315f706c8","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2502.10498","target":"ad7614ff9d335520b4e20c7315f706c8","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"c7b76e0093985d3f8b914a924c3add9c","target":"concept_t0","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2310.08864","target":"c7b76e0093985d3f8b914a924c3add9c","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2506.09985","target":"c7b76e0093985d3f8b914a924c3add9c","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2506.18897","target":"c7b76e0093985d3f8b914a924c3add9c","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2506.21539","target":"c7b76e0093985d3f8b914a924c3add9c","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2603.18532","target":"c7b76e0093985d3f8b914a924c3add9c","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"95bbe72473cf515aa6679ad38582b671","target":"concept_t8","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1706.03741","target":"95bbe72473cf515aa6679ad38582b671","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2305.18153","target":"95bbe72473cf515aa6679ad38582b671","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2402.01695","target":"95bbe72473cf515aa6679ad38582b671","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2406.11944","target":"95bbe72473cf515aa6679ad38582b671","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2407.08867","target":"95bbe72473cf515aa6679ad38582b671","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"362ca2f019995b339e20389e00590b60","target":"concept_t19","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2308.03688","target":"362ca2f019995b339e20389e00590b60","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2402.10196","target":"362ca2f019995b339e20389e00590b60","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2403.16971","target":"362ca2f019995b339e20389e00590b60","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2411.06559","target":"362ca2f019995b339e20389e00590b60","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2601.01885","target":"362ca2f019995b339e20389e00590b60","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"c88a7a9d1b07512998b157da626ea698","target":"concept_t30","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2110.04544","target":"c88a7a9d1b07512998b157da626ea698","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2406.06973","target":"c88a7a9d1b07512998b157da626ea698","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2512.21577","target":"c88a7a9d1b07512998b157da626ea698","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"67bf43b7f5555ac8a73a70ffcc0c7ffc","target":"concept_t1","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2306.05685","target":"67bf43b7f5555ac8a73a70ffcc0c7ffc","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2309.07870","target":"67bf43b7f5555ac8a73a70ffcc0c7ffc","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2312.03815","target":"67bf43b7f5555ac8a73a70ffcc0c7ffc","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2401.02954","target":"67bf43b7f5555ac8a73a70ffcc0c7ffc","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2401.03804","target":"67bf43b7f5555ac8a73a70ffcc0c7ffc","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"844313ebbc105d1b83be279baab0e333","target":"concept_t16","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2305.17926","target":"844313ebbc105d1b83be279baab0e333","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2305.18153","target":"844313ebbc105d1b83be279baab0e333","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2308.12284","target":"844313ebbc105d1b83be279baab0e333","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2309.12288","target":"844313ebbc105d1b83be279baab0e333","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2401.07851","target":"844313ebbc105d1b83be279baab0e333","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"e017307e8bb95c4cb42008a204189377","target":"concept_t19","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1701.06970","target":"e017307e8bb95c4cb42008a204189377","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2405.06211","target":"e017307e8bb95c4cb42008a204189377","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2411.15594","target":"e017307e8bb95c4cb42008a204189377","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2412.14161","target":"e017307e8bb95c4cb42008a204189377","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2502.21321","target":"e017307e8bb95c4cb42008a204189377","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"0cf269af8e0050349e9fa97c24b879f8","target":"concept_t12","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2106.02039","target":"0cf269af8e0050349e9fa97c24b879f8","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2206.04114","target":"0cf269af8e0050349e9fa97c24b879f8","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2206.14176","target":"0cf269af8e0050349e9fa97c24b879f8","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2308.01399","target":"0cf269af8e0050349e9fa97c24b879f8","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2310.06114","target":"0cf269af8e0050349e9fa97c24b879f8","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"cda7dd561a0953c89d239b7026e2b39d","target":"concept_t26","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1811.04551","target":"cda7dd561a0953c89d239b7026e2b39d","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1911.10601","target":"cda7dd561a0953c89d239b7026e2b39d","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1912.01603","target":"cda7dd561a0953c89d239b7026e2b39d","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2205.06175","target":"cda7dd561a0953c89d239b7026e2b39d","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2305.17198","target":"cda7dd561a0953c89d239b7026e2b39d","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"56556af84b9d5c8382bad825092997e0","target":"concept_t30","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2205.06175","target":"56556af84b9d5c8382bad825092997e0","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2303.03378","target":"56556af84b9d5c8382bad825092997e0","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2306.09265","target":"56556af84b9d5c8382bad825092997e0","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2309.02591","target":"56556af84b9d5c8382bad825092997e0","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"f4c0d2e5117b5844aa11a1dd0e8b79c5","target":"concept_t28","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1712.04602","target":"f4c0d2e5117b5844aa11a1dd0e8b79c5","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1805.07526","target":"f4c0d2e5117b5844aa11a1dd0e8b79c5","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2006.04182","target":"f4c0d2e5117b5844aa11a1dd0e8b79c5","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2103.03725","target":"f4c0d2e5117b5844aa11a1dd0e8b79c5","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2207.12316","target":"f4c0d2e5117b5844aa11a1dd0e8b79c5","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"463f8662c60d5656b5416579617882ef","target":"concept_t31","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2403.14403","target":"463f8662c60d5656b5416579617882ef","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2407.02485","target":"463f8662c60d5656b5416579617882ef","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"3f72bc264cb352c986394fe73e6e5d14","target":"concept_t27","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2305.14992","target":"3f72bc264cb352c986394fe73e6e5d14","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2502.03373","target":"3f72bc264cb352c986394fe73e6e5d14","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2502.18080","target":"3f72bc264cb352c986394fe73e6e5d14","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2502.21321","target":"3f72bc264cb352c986394fe73e6e5d14","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2504.10903","target":"3f72bc264cb352c986394fe73e6e5d14","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"ebf143af89b85b7baba9b1c7f66524e9","target":"concept_t32","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2303.01469","target":"ebf143af89b85b7baba9b1c7f66524e9","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2309.02591","target":"ebf143af89b85b7baba9b1c7f66524e9","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2311.05348","target":"ebf143af89b85b7baba9b1c7f66524e9","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2402.12451","target":"ebf143af89b85b7baba9b1c7f66524e9","relation":"kg_evaluated_on","confidence":"EXTRACTED","confidence_score":0.6},{"source":"a0a77dd812b6582d9a29b5dd65c8ace3","target":"concept_t23","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2105.04906","target":"a0a77dd812b6582d9a29b5dd65c8ace3","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2306.02572","target":"a0a77dd812b6582d9a29b5dd65c8ace3","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2307.12698","target":"a0a77dd812b6582d9a29b5dd65c8ace3","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2403.00504","target":"a0a77dd812b6582d9a29b5dd65c8ace3","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2404.08471","target":"a0a77dd812b6582d9a29b5dd65c8ace3","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"53b1c61910fe500e9bce24ac49a31360","target":"concept_t16","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1810.04805","target":"53b1c61910fe500e9bce24ac49a31360","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1910.10683","target":"53b1c61910fe500e9bce24ac49a31360","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2005.14165","target":"53b1c61910fe500e9bce24ac49a31360","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2101.03961","target":"53b1c61910fe500e9bce24ac49a31360","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2305.18654","target":"53b1c61910fe500e9bce24ac49a31360","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"b8667390705e55ac9928e8cf9a645132","target":"concept_t23","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1312.6114","target":"b8667390705e55ac9928e8cf9a645132","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1810.04805","target":"b8667390705e55ac9928e8cf9a645132","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2105.04906","target":"b8667390705e55ac9928e8cf9a645132","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2307.12698","target":"b8667390705e55ac9928e8cf9a645132","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2403.00504","target":"b8667390705e55ac9928e8cf9a645132","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"c3ad4df1e789548683b4021361c66d6b","target":"concept_t32","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2112.10752","target":"c3ad4df1e789548683b4021361c66d6b","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2210.04133","target":"c3ad4df1e789548683b4021361c66d6b","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2406.02347","target":"c3ad4df1e789548683b4021361c66d6b","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"lumalabs.ai:cf7b765e8f15","target":"c3ad4df1e789548683b4021361c66d6b","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"60faede8ef095b4f9e15856e8dec4bec","target":"concept_t0","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2206.11795","target":"60faede8ef095b4f9e15856e8dec4bec","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2210.07729","target":"60faede8ef095b4f9e15856e8dec4bec","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2212.06817","target":"60faede8ef095b4f9e15856e8dec4bec","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"f4fc68cec2a6556792d727805d687e89","target":"concept_t16","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2202.08906","target":"f4fc68cec2a6556792d727805d687e89","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"b8c5abc6079955949e89e8e5737f65cc","target":"concept_t37","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2308.09723","target":"b8c5abc6079955949e89e8e5737f65cc","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2312.03788","target":"b8c5abc6079955949e89e8e5737f65cc","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2410.13056","target":"b8c5abc6079955949e89e8e5737f65cc","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2504.05352","target":"b8c5abc6079955949e89e8e5737f65cc","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"dfbfd28709bd5ca4b61587004947d0fc","target":"concept_t25","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1705.09156","target":"dfbfd28709bd5ca4b61587004947d0fc","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1906.10184","target":"dfbfd28709bd5ca4b61587004947d0fc","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2105.11203","target":"dfbfd28709bd5ca4b61587004947d0fc","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2201.06387","target":"dfbfd28709bd5ca4b61587004947d0fc","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2508.12027","target":"dfbfd28709bd5ca4b61587004947d0fc","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"f723aad37a3d5add8b6b8c2871979688","target":"concept_t20","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2510.20416","target":"f723aad37a3d5add8b6b8c2871979688","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"43c111a7b6d85e62b33bf10e9ab3cd86","target":"concept_t13","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2210.04133","target":"43c111a7b6d85e62b33bf10e9ab3cd86","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2512.00597","target":"43c111a7b6d85e62b33bf10e9ab3cd86","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"5aa08fff1eb55ba9ad84ee93a2cc7d04","target":"concept_t11","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2304.12995","target":"5aa08fff1eb55ba9ad84ee93a2cc7d04","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"a0eb5cd6f4f955ef910072df8960f881","target":"concept_t13","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2407.06508","target":"a0eb5cd6f4f955ef910072df8960f881","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2411.04118","target":"a0eb5cd6f4f955ef910072df8960f881","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2506.22567","target":"a0eb5cd6f4f955ef910072df8960f881","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2507.05201","target":"a0eb5cd6f4f955ef910072df8960f881","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2603.00479","target":"a0eb5cd6f4f955ef910072df8960f881","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"dd3ae0bae15e5514884bf73ca72151e5","target":"concept_t33","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2405.15383","target":"dd3ae0bae15e5514884bf73ca72151e5","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"e520bbf665c753528a6181ff6d7ca2a0","target":"concept_t20","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1705.09156","target":"e520bbf665c753528a6181ff6d7ca2a0","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:1805.07526","target":"e520bbf665c753528a6181ff6d7ca2a0","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2202.09467","target":"e520bbf665c753528a6181ff6d7ca2a0","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"be97f57d3c57559a855c8ed7dd83f07b","target":"concept_t6","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"proceedings.mlr.press:6beb908867f9","target":"be97f57d3c57559a855c8ed7dd83f07b","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"d6c30301190553349449581c5b70071b","target":"concept_t13","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2603.00479","target":"d6c30301190553349449581c5b70071b","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"576e8eab29425f3db82f2837e1afa170","target":"concept_t23","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2105.04906","target":"576e8eab29425f3db82f2837e1afa170","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2309.05767","target":"576e8eab29425f3db82f2837e1afa170","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"d3d1df5c9d765256b66104fd1ad76d1d","target":"concept_t13","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"f4282cf7e07b533da381c9a327992940","target":"concept_t17","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"f4282cf7e07b533da381c9a327992940","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"2547d0055bdf590fbf512b575f47493e","target":"concept_t27","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2305.18654","target":"2547d0055bdf590fbf512b575f47493e","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"transformer-circuits.pub:c4d9a1f0e2b7","target":"2547d0055bdf590fbf512b575f47493e","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"e2cd7f846bb65742a94c810071a398ef","target":"concept_t31","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2312.10997","target":"e2cd7f846bb65742a94c810071a398ef","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2405.06211","target":"e2cd7f846bb65742a94c810071a398ef","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"12dc907108be5ba6bf916312e91379d8","target":"concept_t22","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1812.00922","target":"12dc907108be5ba6bf916312e91379d8","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2202.03771","target":"12dc907108be5ba6bf916312e91379d8","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2606.10749","target":"12dc907108be5ba6bf916312e91379d8","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"94807cc518245d45b6d9b722fdc49931","target":"concept_t22","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"8c19be258e0659bfa040d9bd16b2b852","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1708.00133","target":"8c19be258e0659bfa040d9bd16b2b852","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2408.14837","target":"8c19be258e0659bfa040d9bd16b2b852","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2505.21996","target":"8c19be258e0659bfa040d9bd16b2b852","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2512.04797","target":"8c19be258e0659bfa040d9bd16b2b852","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"ce96547a8bd75f2c950f698f9e0b42d7","target":"concept_t17","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"ce96547a8bd75f2c950f698f9e0b42d7","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"f4da11522c2a5573a1d43f2f81e7df3a","target":"concept_t29","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"f5d42622e0125527b372433a7ec6a846","target":"concept_t11","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1312.6114","target":"f5d42622e0125527b372433a7ec6a846","relation":"kg_about","confidence":"EXTRACTED","confidence_score":0.6},{"source":"e9dea714730e5a2dae3c88f69887a31c","target":"concept_t17","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"e9dea714730e5a2dae3c88f69887a31c","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"ee22adec46ca5d95a110b38805bb1097","target":"concept_t17","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"ee22adec46ca5d95a110b38805bb1097","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"dc86e7812f445fa6a43e21a83dc1b287","target":"concept_t17","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"dc86e7812f445fa6a43e21a83dc1b287","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"d49321f4b3105e6f9f83a04623bad056","target":"concept_t17","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"d49321f4b3105e6f9f83a04623bad056","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"c6594b9752b051f58bc98e3c76a46cad","target":"concept_t17","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"c6594b9752b051f58bc98e3c76a46cad","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"cd4ebce009bc570da4a44f22bbdf9ba8","target":"concept_t17","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"cd4ebce009bc570da4a44f22bbdf9ba8","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"0c7c7b4eb4b5569dbff6d810e7784c49","target":"concept_t23","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2304.07193","target":"0c7c7b4eb4b5569dbff6d810e7784c49","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"5ab80843a1a1566ebb7cd7bdb5120e5d","target":"concept_t17","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"5ab80843a1a1566ebb7cd7bdb5120e5d","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"35a196c11652577b9801ae4fefc157d0","target":"concept_t17","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"1x.tech:d48ab282348b","target":"35a196c11652577b9801ae4fefc157d0","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"4068f7c4e1f6531daba0ce4f4275a89b","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"thesequence.substack.com:0f25bda15c4b","target":"4068f7c4e1f6531daba0ce4f4275a89b","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"637df3c30efd5f9f91bc799b491146f1","target":"concept_t19","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2303.11366","target":"637df3c30efd5f9f91bc799b491146f1","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2305.16653","target":"637df3c30efd5f9f91bc799b491146f1","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2310.04276","target":"637df3c30efd5f9f91bc799b491146f1","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"c9df46f9eaf15c4d9c3a5ce01a64c1b7","target":"concept_t0","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2310.01361","target":"c9df46f9eaf15c4d9c3a5ce01a64c1b7","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"6318ed26b8415943b81b6e73282339cc","target":"concept_t19","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2303.11366","target":"6318ed26b8415943b81b6e73282339cc","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2304.03442","target":"6318ed26b8415943b81b6e73282339cc","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"486baa7031875dadadd3ee7d83a7a94d","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2306.02572","target":"486baa7031875dadadd3ee7d83a7a94d","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"openreview.net:3f48a33ff9bf","target":"486baa7031875dadadd3ee7d83a7a94d","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"e64a473ca88450d0bd31bf1a8b7118e3","target":"concept_t3","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1312.6114","target":"e64a473ca88450d0bd31bf1a8b7118e3","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"ef539d7d7d345e4d816f27b3a4120bcb","target":"concept_t26","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1910.08210","target":"ef539d7d7d345e4d816f27b3a4120bcb","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2305.05383","target":"ef539d7d7d345e4d816f27b3a4120bcb","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"24c813c93c21563b8f444f5c8a02ad89","target":"concept_t14","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2306.02572","target":"24c813c93c21563b8f444f5c8a02ad89","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"77064b78bc7458b28cf78b7f3e5eeb73","target":"concept_t19","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2305.16653","target":"77064b78bc7458b28cf78b7f3e5eeb73","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"a65e041dcf2e5692a4e593140113a16d","target":"concept_t23","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2111.06377","target":"a65e041dcf2e5692a4e593140113a16d","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2112.10752","target":"a65e041dcf2e5692a4e593140113a16d","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"cc066218775353349cac77b5cf3091a3","target":"concept_t1","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1910.10683","target":"cc066218775353349cac77b5cf3091a3","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2005.14165","target":"cc066218775353349cac77b5cf3091a3","relation":"kg_mentions","confidence":"EXTRACTED","confidence_score":0.6},{"source":"817a3ffd1c285c55b8a2eb938ba85f4e","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"oasis-model.github.io:525053739707","target":"817a3ffd1c285c55b8a2eb938ba85f4e","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"c2c0e7d5306a5897b8bc22c8fce99645","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"deepmind.google:f020e8dd5471","target":"c2c0e7d5306a5897b8bc22c8fce99645","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"d8e9b19f33135d92b63f857a31154224","target":"concept_t0","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2410.24164","target":"d8e9b19f33135d92b63f857a31154224","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"1106a1b9f3515e9290d700aa9e21f35b","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"5547843881175855b026a2df0af18c7a","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"5e29463a898b5d3ab547102ee9ce6bfd","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"github.com:14ee211b0b5d","target":"5e29463a898b5d3ab547102ee9ce6bfd","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"addc77b5856b50cd8a6642888e8d8b96","target":"concept_t30","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"b857f1ae5f585f12bed9b48d888ca7ab","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2506.09985","target":"b857f1ae5f585f12bed9b48d888ca7ab","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"cb06a17fae925dbeb50d75fd3daf1b1c","target":"concept_t0","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"26a42a6a9311514da6bbc37bba8457ec","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"8d34e1fc3d105707a21910c27e565b82","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"9df51276f50f503db60d7f16830fbc6b","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"c483220c93825f1bab9625e6d8aecd9f","target":"concept_t25","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"nature.com:7fc0dd8e4038","target":"c483220c93825f1bab9625e6d8aecd9f","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"ec00eb3c034e5985bdb7ddf4866f493e","target":"concept_t17","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"f11c75d47c2e53a4b94fc0a39810c80e","target":"concept_t32","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2402.15391","target":"f11c75d47c2e53a4b94fc0a39810c80e","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"142d56d7638b53659e24beadb82517fb","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"deepmind.google:f020e8dd5471","target":"142d56d7638b53659e24beadb82517fb","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"1499a1f59c7c5b938b63a596dce80547","target":"concept_t8","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"167f7f0675155167b1407dbd05df9210","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"2265ce6d85fa5f579da9f46b6178054a","target":"concept_t29","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"37bd6a5426ca5981bb60f3765730bd51","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"425e1e76829350378eb6218d132f8d01","target":"concept_t19","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"5b9fd3728a5a5cc6a157247791175001","target":"concept_t0","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"715d0b7b2c4f56d4a0d8c9be9684c5e6","target":"concept_t12","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"nianticlabs.com:05a236f27efe","target":"715d0b7b2c4f56d4a0d8c9be9684c5e6","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"8669e2f0da50530491a3851cadce7e1a","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"8de0b6d5720051d8bc1d8de5781c3155","target":"concept_t25","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:1906.10184","target":"8de0b6d5720051d8bc1d8de5781c3155","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2002.04501","target":"8de0b6d5720051d8bc1d8de5781c3155","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"983982f2750e51e2ac18870a3e1238d7","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"bf0453e0344b54c99629b05d71d88721","target":"concept_t20","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2105.07308","target":"bf0453e0344b54c99629b05d71d88721","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"arxiv:2308.07870","target":"bf0453e0344b54c99629b05d71d88721","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"e9707c46872b54fc8d0ab6898b5a8b66","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"ef96446cbb435bdc87f8701e55712374","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"efc5d4f358ce5894b78284ab94cc7bda","target":"concept_t30","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"f633af356dbf5f378271c0dc2800d089","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"02e907aa5b48500e87e03a1cd9b4c138","target":"concept_t19","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"02f7fb51dd405cf7a36ebcdf961d23fd","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"0bf8bc0d26195a2da57b02e5fc304610","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"0d964edb4353516da96e5aee3d389c8c","target":"concept_t0","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2402.15391","target":"0d964edb4353516da96e5aee3d389c8c","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"0fb2127e562b59e5893fc9889bcb5b15","target":"concept_t35","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"arxiv:2507.12440","target":"0fb2127e562b59e5893fc9889bcb5b15","relation":"kg_introduces","confidence":"EXTRACTED","confidence_score":0.6},{"source":"135ad2f14b9a517fae601d9d928531c3","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"175e9a9b50aa586bac9fc5c9eb0aec22","target":"concept_t38","relation":"kg-anchor","confidence":"EXTRACTED","confidence_score":0.5},{"source":"0a7193294bbb5d1f9097552f2242f34f","target":"ef9d9f01afc951daaba0da0ceaeb4988","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.393},{"source":"0a7193294bbb5d1f9097552f2242f34f","target":"b76fb931c1c75ba7964ec119fc2a0b34","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.305},{"source":"0a7193294bbb5d1f9097552f2242f34f","target":"e2e9917abed652b59f2892d75ccdec63","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.299},{"source":"0a7193294bbb5d1f9097552f2242f34f","target":"3309b4841a9456de8a44882428576753","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.268},{"source":"0a7193294bbb5d1f9097552f2242f34f","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.233},{"source":"0a7193294bbb5d1f9097552f2242f34f","target":"00ed333d471b517cacd77ef474592c37","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.231},{"source":"c5c0a185adae51f49af390e3d5f679c6","target":"1ea0320008f054d6a48355261be0862a","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.694},{"source":"c5c0a185adae51f49af390e3d5f679c6","target":"578297fe62465056b5d74185e59b4615","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.67},{"source":"c5c0a185adae51f49af390e3d5f679c6","target":"00ed333d471b517cacd77ef474592c37","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.409},{"source":"c5c0a185adae51f49af390e3d5f679c6","target":"5e88711b4de2546296d173f144a8ca38","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.399},{"source":"c5c0a185adae51f49af390e3d5f679c6","target":"1fbb24e7e94e58e18899bd2b06c8fed4","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.383},{"source":"c5c0a185adae51f49af390e3d5f679c6","target":"d5699c11ff3e5f2ea251177377b18fff","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.341},{"source":"00ed333d471b517cacd77ef474592c37","target":"6aeb0ae745b350e18820468d8ce4df3e","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.444},{"source":"00ed333d471b517cacd77ef474592c37","target":"1ea0320008f054d6a48355261be0862a","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.368},{"source":"00ed333d471b517cacd77ef474592c37","target":"578297fe62465056b5d74185e59b4615","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.343},{"source":"00ed333d471b517cacd77ef474592c37","target":"1fbb24e7e94e58e18899bd2b06c8fed4","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.303},{"source":"00ed333d471b517cacd77ef474592c37","target":"5e88711b4de2546296d173f144a8ca38","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.293},{"source":"1ea0320008f054d6a48355261be0862a","target":"578297fe62465056b5d74185e59b4615","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.597},{"source":"1ea0320008f054d6a48355261be0862a","target":"d5699c11ff3e5f2ea251177377b18fff","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.397},{"source":"1ea0320008f054d6a48355261be0862a","target":"5e88711b4de2546296d173f144a8ca38","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.373},{"source":"1ea0320008f054d6a48355261be0862a","target":"1fbb24e7e94e58e18899bd2b06c8fed4","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.347},{"source":"e2e9917abed652b59f2892d75ccdec63","target":"00ed333d471b517cacd77ef474592c37","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.212},{"source":"e2e9917abed652b59f2892d75ccdec63","target":"afbadd921ac75a03a7cb4ed228efc708","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.198},{"source":"e2e9917abed652b59f2892d75ccdec63","target":"cf4ebe8a6fd0551fa06e22cccb6525d9","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.173},{"source":"e2e9917abed652b59f2892d75ccdec63","target":"b76fb931c1c75ba7964ec119fc2a0b34","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.168},{"source":"e2e9917abed652b59f2892d75ccdec63","target":"ef9d9f01afc951daaba0da0ceaeb4988","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.167},{"source":"ef9d9f01afc951daaba0da0ceaeb4988","target":"b76fb931c1c75ba7964ec119fc2a0b34","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.277},{"source":"ef9d9f01afc951daaba0da0ceaeb4988","target":"3309b4841a9456de8a44882428576753","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.256},{"source":"ef9d9f01afc951daaba0da0ceaeb4988","target":"1fbb24e7e94e58e18899bd2b06c8fed4","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.167},{"source":"ef9d9f01afc951daaba0da0ceaeb4988","target":"1ea0320008f054d6a48355261be0862a","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.156},{"source":"5e88711b4de2546296d173f144a8ca38","target":"3e5f643d839e5aa2b8723a46fba6a2a6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.362},{"source":"5e88711b4de2546296d173f144a8ca38","target":"1fbb24e7e94e58e18899bd2b06c8fed4","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.336},{"source":"5e88711b4de2546296d173f144a8ca38","target":"d613731276125291ae874322d754e3e0","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.325},{"source":"5e88711b4de2546296d173f144a8ca38","target":"578297fe62465056b5d74185e59b4615","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.319},{"source":"6aeb0ae745b350e18820468d8ce4df3e","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.338},{"source":"6aeb0ae745b350e18820468d8ce4df3e","target":"1ea0320008f054d6a48355261be0862a","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.285},{"source":"6aeb0ae745b350e18820468d8ce4df3e","target":"578297fe62465056b5d74185e59b4615","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.277},{"source":"6aeb0ae745b350e18820468d8ce4df3e","target":"5e88711b4de2546296d173f144a8ca38","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.27},{"source":"6aeb0ae745b350e18820468d8ce4df3e","target":"2fe319dfa7aa5c34b3373b94d57bfacc","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.243},{"source":"1fbb24e7e94e58e18899bd2b06c8fed4","target":"578297fe62465056b5d74185e59b4615","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.324},{"source":"1fbb24e7e94e58e18899bd2b06c8fed4","target":"d5699c11ff3e5f2ea251177377b18fff","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.245},{"source":"b76fb931c1c75ba7964ec119fc2a0b34","target":"3309b4841a9456de8a44882428576753","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.431},{"source":"b76fb931c1c75ba7964ec119fc2a0b34","target":"5d706c4d8e3b5c51bf1534e0ac3649be","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.144},{"source":"b76fb931c1c75ba7964ec119fc2a0b34","target":"1fbb24e7e94e58e18899bd2b06c8fed4","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.134},{"source":"578297fe62465056b5d74185e59b4615","target":"179eca3284ca53d1909a4865bb5c5f67","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.331},{"source":"578297fe62465056b5d74185e59b4615","target":"d5699c11ff3e5f2ea251177377b18fff","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.326},{"source":"b091ad18c2395679bfef4cda988e8378","target":"3e5f643d839e5aa2b8723a46fba6a2a6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.51},{"source":"b091ad18c2395679bfef4cda988e8378","target":"d613731276125291ae874322d754e3e0","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.444},{"source":"b091ad18c2395679bfef4cda988e8378","target":"5e88711b4de2546296d173f144a8ca38","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.304},{"source":"b091ad18c2395679bfef4cda988e8378","target":"f7d9b5d7ab0e51f8a085ffe40543f902","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.233},{"source":"b091ad18c2395679bfef4cda988e8378","target":"1ea0320008f054d6a48355261be0862a","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.197},{"source":"b091ad18c2395679bfef4cda988e8378","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.195},{"source":"3309b4841a9456de8a44882428576753","target":"1ea0320008f054d6a48355261be0862a","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.188},{"source":"3309b4841a9456de8a44882428576753","target":"1fbb24e7e94e58e18899bd2b06c8fed4","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.174},{"source":"3309b4841a9456de8a44882428576753","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.172},{"source":"3e5f643d839e5aa2b8723a46fba6a2a6","target":"d613731276125291ae874322d754e3e0","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.631},{"source":"3e5f643d839e5aa2b8723a46fba6a2a6","target":"f7d9b5d7ab0e51f8a085ffe40543f902","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.233},{"source":"3e5f643d839e5aa2b8723a46fba6a2a6","target":"1ea0320008f054d6a48355261be0862a","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.222},{"source":"3e5f643d839e5aa2b8723a46fba6a2a6","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.204},{"source":"cf4ebe8a6fd0551fa06e22cccb6525d9","target":"5a9f7bc5b9f45daca6a11a434b3f61d9","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.192},{"source":"cf4ebe8a6fd0551fa06e22cccb6525d9","target":"00ed333d471b517cacd77ef474592c37","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.148},{"source":"cf4ebe8a6fd0551fa06e22cccb6525d9","target":"0a7193294bbb5d1f9097552f2242f34f","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.138},{"source":"cf4ebe8a6fd0551fa06e22cccb6525d9","target":"afbadd921ac75a03a7cb4ed228efc708","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.122},{"source":"cf4ebe8a6fd0551fa06e22cccb6525d9","target":"6aeb0ae745b350e18820468d8ce4df3e","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.121},{"source":"afbadd921ac75a03a7cb4ed228efc708","target":"55dc64e751fc5cada055cbef1ba6f490","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.138},{"source":"afbadd921ac75a03a7cb4ed228efc708","target":"5a237ad74e26549290d5b57774b015da","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.129},{"source":"afbadd921ac75a03a7cb4ed228efc708","target":"29e518447cab5ffeab46a89a0e892ece","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.127},{"source":"afbadd921ac75a03a7cb4ed228efc708","target":"0a7193294bbb5d1f9097552f2242f34f","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.125},{"source":"5d706c4d8e3b5c51bf1534e0ac3649be","target":"e0312e04abde5e53a1f0b3a14d1422ba","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.179},{"source":"5d706c4d8e3b5c51bf1534e0ac3649be","target":"5a9f7bc5b9f45daca6a11a434b3f61d9","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.165},{"source":"5d706c4d8e3b5c51bf1534e0ac3649be","target":"0a7193294bbb5d1f9097552f2242f34f","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.134},{"source":"5d706c4d8e3b5c51bf1534e0ac3649be","target":"e2e9917abed652b59f2892d75ccdec63","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.104},{"source":"5d706c4d8e3b5c51bf1534e0ac3649be","target":"3309b4841a9456de8a44882428576753","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.101},{"source":"d613731276125291ae874322d754e3e0","target":"f7d9b5d7ab0e51f8a085ffe40543f902","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.238},{"source":"d613731276125291ae874322d754e3e0","target":"0bf069239279577ebdb04c2e7f09c1c9","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.208},{"source":"d613731276125291ae874322d754e3e0","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.191},{"source":"f7d9b5d7ab0e51f8a085ffe40543f902","target":"5e88711b4de2546296d173f144a8ca38","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.229},{"source":"f7d9b5d7ab0e51f8a085ffe40543f902","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.227},{"source":"f7d9b5d7ab0e51f8a085ffe40543f902","target":"1ea0320008f054d6a48355261be0862a","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.227},{"source":"d5699c11ff3e5f2ea251177377b18fff","target":"179eca3284ca53d1909a4865bb5c5f67","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.281},{"source":"d5699c11ff3e5f2ea251177377b18fff","target":"5e88711b4de2546296d173f144a8ca38","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.209},{"source":"e0312e04abde5e53a1f0b3a14d1422ba","target":"e2e9917abed652b59f2892d75ccdec63","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.159},{"source":"e0312e04abde5e53a1f0b3a14d1422ba","target":"95bbe72473cf515aa6679ad38582b671","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.14},{"source":"e0312e04abde5e53a1f0b3a14d1422ba","target":"0a7193294bbb5d1f9097552f2242f34f","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.116},{"source":"e0312e04abde5e53a1f0b3a14d1422ba","target":"b76fb931c1c75ba7964ec119fc2a0b34","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.105},{"source":"e0312e04abde5e53a1f0b3a14d1422ba","target":"41a9b2d7e67f56009ef5f9b25218ae23","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.103},{"source":"0bf069239279577ebdb04c2e7f09c1c9","target":"156bd9a2ef6c50b78065af31965a80e8","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.57},{"source":"0bf069239279577ebdb04c2e7f09c1c9","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.25},{"source":"0bf069239279577ebdb04c2e7f09c1c9","target":"f7d9b5d7ab0e51f8a085ffe40543f902","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.214},{"source":"0bf069239279577ebdb04c2e7f09c1c9","target":"1ea0320008f054d6a48355261be0862a","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.208},{"source":"0bf069239279577ebdb04c2e7f09c1c9","target":"5e88711b4de2546296d173f144a8ca38","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.202},{"source":"5a9f7bc5b9f45daca6a11a434b3f61d9","target":"20e90bf1db27548698a1a21ebe2a12da","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.436},{"source":"5a9f7bc5b9f45daca6a11a434b3f61d9","target":"5e88711b4de2546296d173f144a8ca38","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.118},{"source":"5a9f7bc5b9f45daca6a11a434b3f61d9","target":"41a9b2d7e67f56009ef5f9b25218ae23","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.111},{"source":"5a9f7bc5b9f45daca6a11a434b3f61d9","target":"f4c0d2e5117b5844aa11a1dd0e8b79c5","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.104},{"source":"2fe319dfa7aa5c34b3373b94d57bfacc","target":"00ed333d471b517cacd77ef474592c37","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.199},{"source":"2fe319dfa7aa5c34b3373b94d57bfacc","target":"5e88711b4de2546296d173f144a8ca38","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.193},{"source":"2fe319dfa7aa5c34b3373b94d57bfacc","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.171},{"source":"2fe319dfa7aa5c34b3373b94d57bfacc","target":"1ea0320008f054d6a48355261be0862a","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.154},{"source":"2fe319dfa7aa5c34b3373b94d57bfacc","target":"da6911ef37435c6abb169e201aa5f45d","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.146},{"source":"179eca3284ca53d1909a4865bb5c5f67","target":"1ea0320008f054d6a48355261be0862a","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.279},{"source":"179eca3284ca53d1909a4865bb5c5f67","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.253},{"source":"179eca3284ca53d1909a4865bb5c5f67","target":"8b61b4feb56658e4877bb944c345a611","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.162},{"source":"179eca3284ca53d1909a4865bb5c5f67","target":"1fbb24e7e94e58e18899bd2b06c8fed4","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.157},{"source":"41a9b2d7e67f56009ef5f9b25218ae23","target":"0bf069239279577ebdb04c2e7f09c1c9","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.127},{"source":"41a9b2d7e67f56009ef5f9b25218ae23","target":"156bd9a2ef6c50b78065af31965a80e8","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.097},{"source":"41a9b2d7e67f56009ef5f9b25218ae23","target":"cf4ebe8a6fd0551fa06e22cccb6525d9","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.092},{"source":"41a9b2d7e67f56009ef5f9b25218ae23","target":"2fe319dfa7aa5c34b3373b94d57bfacc","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.086},{"source":"55dc64e751fc5cada055cbef1ba6f490","target":"e2e9917abed652b59f2892d75ccdec63","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.14},{"source":"55dc64e751fc5cada055cbef1ba6f490","target":"6cb325eebf8b5cf1b32fd77c46351199","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.119},{"source":"55dc64e751fc5cada055cbef1ba6f490","target":"362ca2f019995b339e20389e00590b60","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.097},{"source":"55dc64e751fc5cada055cbef1ba6f490","target":"5a237ad74e26549290d5b57774b015da","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.08},{"source":"55dc64e751fc5cada055cbef1ba6f490","target":"3abafab4d0695bef9da5342372fe06ce","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.072},{"source":"156bd9a2ef6c50b78065af31965a80e8","target":"f7d9b5d7ab0e51f8a085ffe40543f902","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.166},{"source":"156bd9a2ef6c50b78065af31965a80e8","target":"d613731276125291ae874322d754e3e0","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.154},{"source":"156bd9a2ef6c50b78065af31965a80e8","target":"5e88711b4de2546296d173f144a8ca38","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.15},{"source":"156bd9a2ef6c50b78065af31965a80e8","target":"3e5f643d839e5aa2b8723a46fba6a2a6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.147},{"source":"156bd9a2ef6c50b78065af31965a80e8","target":"578297fe62465056b5d74185e59b4615","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.145},{"source":"20e90bf1db27548698a1a21ebe2a12da","target":"cf4ebe8a6fd0551fa06e22cccb6525d9","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.12},{"source":"20e90bf1db27548698a1a21ebe2a12da","target":"5d706c4d8e3b5c51bf1534e0ac3649be","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.098},{"source":"20e90bf1db27548698a1a21ebe2a12da","target":"f4c0d2e5117b5844aa11a1dd0e8b79c5","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.076},{"source":"20e90bf1db27548698a1a21ebe2a12da","target":"41a9b2d7e67f56009ef5f9b25218ae23","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.074},{"source":"dfdc4dcc97d55b138155289efc1f8f15","target":"b091ad18c2395679bfef4cda988e8378","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.18},{"source":"dfdc4dcc97d55b138155289efc1f8f15","target":"ebf143af89b85b7baba9b1c7f66524e9","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.13},{"source":"dfdc4dcc97d55b138155289efc1f8f15","target":"3e5f643d839e5aa2b8723a46fba6a2a6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.121},{"source":"dfdc4dcc97d55b138155289efc1f8f15","target":"c3ad4df1e789548683b4021361c66d6b","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.12},{"source":"dfdc4dcc97d55b138155289efc1f8f15","target":"d613731276125291ae874322d754e3e0","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.086},{"source":"6efe5430e683589693361dbeaf1aa256","target":"afbadd921ac75a03a7cb4ed228efc708","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.09},{"source":"6efe5430e683589693361dbeaf1aa256","target":"29e518447cab5ffeab46a89a0e892ece","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.082},{"source":"6efe5430e683589693361dbeaf1aa256","target":"e2e9917abed652b59f2892d75ccdec63","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.079},{"source":"6efe5430e683589693361dbeaf1aa256","target":"5d706c4d8e3b5c51bf1534e0ac3649be","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.078},{"source":"6efe5430e683589693361dbeaf1aa256","target":"53b1c61910fe500e9bce24ac49a31360","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.075},{"source":"3abafab4d0695bef9da5342372fe06ce","target":"25f5883a505b5c90b3288786468e427b","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.176},{"source":"3abafab4d0695bef9da5342372fe06ce","target":"6cb325eebf8b5cf1b32fd77c46351199","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.158},{"source":"3abafab4d0695bef9da5342372fe06ce","target":"e2e9917abed652b59f2892d75ccdec63","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.103},{"source":"3abafab4d0695bef9da5342372fe06ce","target":"3f72bc264cb352c986394fe73e6e5d14","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.1},{"source":"3abafab4d0695bef9da5342372fe06ce","target":"afbadd921ac75a03a7cb4ed228efc708","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.088},{"source":"3abafab4d0695bef9da5342372fe06ce","target":"dd3ae0bae15e5514884bf73ca72151e5","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.078},{"source":"da6911ef37435c6abb169e201aa5f45d","target":"6aeb0ae745b350e18820468d8ce4df3e","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.137},{"source":"da6911ef37435c6abb169e201aa5f45d","target":"cda7dd561a0953c89d239b7026e2b39d","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.12},{"source":"da6911ef37435c6abb169e201aa5f45d","target":"00ed333d471b517cacd77ef474592c37","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.084},{"source":"da6911ef37435c6abb169e201aa5f45d","target":"5e88711b4de2546296d173f144a8ca38","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.081},{"source":"da6911ef37435c6abb169e201aa5f45d","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.072},{"source":"6f9da474895254f89a0b8c2c1d528480","target":"a0eb5cd6f4f955ef910072df8960f881","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.086},{"source":"6f9da474895254f89a0b8c2c1d528480","target":"d8c361b6739f56a582950c29041b3153","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.084},{"source":"6f9da474895254f89a0b8c2c1d528480","target":"41a9b2d7e67f56009ef5f9b25218ae23","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.083},{"source":"1320c285fbd7589ea3b5acd47bdc24e5","target":"463f8662c60d5656b5416579617882ef","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.149},{"source":"1320c285fbd7589ea3b5acd47bdc24e5","target":"e2e9917abed652b59f2892d75ccdec63","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.085},{"source":"a4a21ac4385e5f19aae2c6647ebddd70","target":"5a9f7bc5b9f45daca6a11a434b3f61d9","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.078},{"source":"a4a21ac4385e5f19aae2c6647ebddd70","target":"f4c0d2e5117b5844aa11a1dd0e8b79c5","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.073},{"source":"de68e0ca85d358759a695e4fe668563e","target":"56556af84b9d5c8382bad825092997e0","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.306},{"source":"de68e0ca85d358759a695e4fe668563e","target":"c88a7a9d1b07512998b157da626ea698","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.152},{"source":"de68e0ca85d358759a695e4fe668563e","target":"3309b4841a9456de8a44882428576753","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.141},{"source":"de68e0ca85d358759a695e4fe668563e","target":"b76fb931c1c75ba7964ec119fc2a0b34","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.11},{"source":"07c2931e095759c7a2114e8d9c086a6b","target":"5aa08fff1eb55ba9ad84ee93a2cc7d04","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.108},{"source":"3a73876784785037a96bc25790a0f5f4","target":"3309b4841a9456de8a44882428576753","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.109},{"source":"3a73876784785037a96bc25790a0f5f4","target":"b76fb931c1c75ba7964ec119fc2a0b34","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.094},{"source":"3a73876784785037a96bc25790a0f5f4","target":"3e5f643d839e5aa2b8723a46fba6a2a6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.081},{"source":"a7af1de80dc55255a1b68d98e0e459d0","target":"f7d9b5d7ab0e51f8a085ffe40543f902","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.193},{"source":"a7af1de80dc55255a1b68d98e0e459d0","target":"156bd9a2ef6c50b78065af31965a80e8","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.144},{"source":"a7af1de80dc55255a1b68d98e0e459d0","target":"d613731276125291ae874322d754e3e0","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.131},{"source":"a7af1de80dc55255a1b68d98e0e459d0","target":"0bf069239279577ebdb04c2e7f09c1c9","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.122},{"source":"a7af1de80dc55255a1b68d98e0e459d0","target":"3e5f643d839e5aa2b8723a46fba6a2a6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.116},{"source":"a7af1de80dc55255a1b68d98e0e459d0","target":"b091ad18c2395679bfef4cda988e8378","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.1},{"source":"e4011f27b2e150bea808160e79f649c6","target":"11e409aa7db05ac6926994cc5efd7a4e","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.187},{"source":"e4011f27b2e150bea808160e79f649c6","target":"a0a77dd812b6582d9a29b5dd65c8ace3","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.09},{"source":"e4011f27b2e150bea808160e79f649c6","target":"5d706c4d8e3b5c51bf1534e0ac3649be","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.083},{"source":"e4011f27b2e150bea808160e79f649c6","target":"b8667390705e55ac9928e8cf9a645132","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.083},{"source":"25f5883a505b5c90b3288786468e427b","target":"29e518447cab5ffeab46a89a0e892ece","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.079},{"source":"11e409aa7db05ac6926994cc5efd7a4e","target":"a0a77dd812b6582d9a29b5dd65c8ace3","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.084},{"source":"11e409aa7db05ac6926994cc5efd7a4e","target":"5d706c4d8e3b5c51bf1534e0ac3649be","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.076},{"source":"11e409aa7db05ac6926994cc5efd7a4e","target":"b8667390705e55ac9928e8cf9a645132","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.073},{"source":"8b61b4feb56658e4877bb944c345a611","target":"578297fe62465056b5d74185e59b4615","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.102},{"source":"8b61b4feb56658e4877bb944c345a611","target":"d5699c11ff3e5f2ea251177377b18fff","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.087},{"source":"8b61b4feb56658e4877bb944c345a611","target":"1ea0320008f054d6a48355261be0862a","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.075},{"source":"8b61b4feb56658e4877bb944c345a611","target":"c5c0a185adae51f49af390e3d5f679c6","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.072},{"source":"d8c361b6739f56a582950c29041b3153","target":"e017307e8bb95c4cb42008a204189377","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.089},{"source":"d8c361b6739f56a582950c29041b3153","target":"41a9b2d7e67f56009ef5f9b25218ae23","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.072},{"source":"d8c361b6739f56a582950c29041b3153","target":"33ec7dab157f55ae97bffe3facf6a915","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.071},{"source":"29e518447cab5ffeab46a89a0e892ece","target":"b8c5abc6079955949e89e8e5737f65cc","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.074},{"source":"ad7614ff9d335520b4e20c7315f706c8","target":"156bd9a2ef6c50b78065af31965a80e8","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.085},{"source":"ad7614ff9d335520b4e20c7315f706c8","target":"a7af1de80dc55255a1b68d98e0e459d0","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.078},{"source":"c7b76e0093985d3f8b914a924c3add9c","target":"179eca3284ca53d1909a4865bb5c5f67","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.149},{"source":"c7b76e0093985d3f8b914a924c3add9c","target":"d5699c11ff3e5f2ea251177377b18fff","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.082},{"source":"c7b76e0093985d3f8b914a924c3add9c","target":"578297fe62465056b5d74185e59b4615","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.077},{"source":"c88a7a9d1b07512998b157da626ea698","target":"56556af84b9d5c8382bad825092997e0","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.144},{"source":"844313ebbc105d1b83be279baab0e333","target":"53b1c61910fe500e9bce24ac49a31360","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.074},{"source":"f4c0d2e5117b5844aa11a1dd0e8b79c5","target":"e520bbf665c753528a6181ff6d7ca2a0","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.08},{"source":"a0a77dd812b6582d9a29b5dd65c8ace3","target":"b8667390705e55ac9928e8cf9a645132","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.394},{"source":"a0a77dd812b6582d9a29b5dd65c8ace3","target":"576e8eab29425f3db82f2837e1afa170","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.164},{"source":"53b1c61910fe500e9bce24ac49a31360","target":"b8c5abc6079955949e89e8e5737f65cc","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.139},{"source":"b8667390705e55ac9928e8cf9a645132","target":"576e8eab29425f3db82f2837e1afa170","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.16},{"source":"dfbfd28709bd5ca4b61587004947d0fc","target":"e520bbf665c753528a6181ff6d7ca2a0","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.1},{"source":"43c111a7b6d85e62b33bf10e9ab3cd86","target":"d6c30301190553349449581c5b70071b","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.183},{"source":"d3d1df5c9d765256b66104fd1ad76d1d","target":"f5d42622e0125527b372433a7ec6a846","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.105},{"source":"f4282cf7e07b533da381c9a327992940","target":"ee22adec46ca5d95a110b38805bb1097","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.8},{"source":"f4282cf7e07b533da381c9a327992940","target":"e9dea714730e5a2dae3c88f69887a31c","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.795},{"source":"f4282cf7e07b533da381c9a327992940","target":"dc86e7812f445fa6a43e21a83dc1b287","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.773},{"source":"f4282cf7e07b533da381c9a327992940","target":"d49321f4b3105e6f9f83a04623bad056","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.75},{"source":"f4282cf7e07b533da381c9a327992940","target":"cd4ebce009bc570da4a44f22bbdf9ba8","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.705},{"source":"f4282cf7e07b533da381c9a327992940","target":"c6594b9752b051f58bc98e3c76a46cad","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.652},{"source":"ce96547a8bd75f2c950f698f9e0b42d7","target":"d49321f4b3105e6f9f83a04623bad056","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.8},{"source":"ce96547a8bd75f2c950f698f9e0b42d7","target":"cd4ebce009bc570da4a44f22bbdf9ba8","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.795},{"source":"ce96547a8bd75f2c950f698f9e0b42d7","target":"dc86e7812f445fa6a43e21a83dc1b287","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.78},{"source":"ce96547a8bd75f2c950f698f9e0b42d7","target":"e9dea714730e5a2dae3c88f69887a31c","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.762},{"source":"ce96547a8bd75f2c950f698f9e0b42d7","target":"c6594b9752b051f58bc98e3c76a46cad","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.732},{"source":"ce96547a8bd75f2c950f698f9e0b42d7","target":"ee22adec46ca5d95a110b38805bb1097","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.727},{"source":"e9dea714730e5a2dae3c88f69887a31c","target":"dc86e7812f445fa6a43e21a83dc1b287","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.971},{"source":"e9dea714730e5a2dae3c88f69887a31c","target":"ee22adec46ca5d95a110b38805bb1097","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.946},{"source":"e9dea714730e5a2dae3c88f69887a31c","target":"d49321f4b3105e6f9f83a04623bad056","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.943},{"source":"e9dea714730e5a2dae3c88f69887a31c","target":"cd4ebce009bc570da4a44f22bbdf9ba8","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.886},{"source":"e9dea714730e5a2dae3c88f69887a31c","target":"c6594b9752b051f58bc98e3c76a46cad","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.811},{"source":"ee22adec46ca5d95a110b38805bb1097","target":"dc86e7812f445fa6a43e21a83dc1b287","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.919},{"source":"ee22adec46ca5d95a110b38805bb1097","target":"d49321f4b3105e6f9f83a04623bad056","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.892},{"source":"ee22adec46ca5d95a110b38805bb1097","target":"cd4ebce009bc570da4a44f22bbdf9ba8","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.838},{"source":"ee22adec46ca5d95a110b38805bb1097","target":"c6594b9752b051f58bc98e3c76a46cad","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.769},{"source":"dc86e7812f445fa6a43e21a83dc1b287","target":"d49321f4b3105e6f9f83a04623bad056","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.971},{"source":"dc86e7812f445fa6a43e21a83dc1b287","target":"cd4ebce009bc570da4a44f22bbdf9ba8","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.912},{"source":"dc86e7812f445fa6a43e21a83dc1b287","target":"c6594b9752b051f58bc98e3c76a46cad","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.833},{"source":"d49321f4b3105e6f9f83a04623bad056","target":"cd4ebce009bc570da4a44f22bbdf9ba8","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.939},{"source":"d49321f4b3105e6f9f83a04623bad056","target":"c6594b9752b051f58bc98e3c76a46cad","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.857},{"source":"c6594b9752b051f58bc98e3c76a46cad","target":"cd4ebce009bc570da4a44f22bbdf9ba8","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.909},{"source":"4068f7c4e1f6531daba0ce4f4275a89b","target":"8669e2f0da50530491a3851cadce7e1a","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.333},{"source":"4068f7c4e1f6531daba0ce4f4275a89b","target":"0bf8bc0d26195a2da57b02e5fc304610","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.118},{"source":"637df3c30efd5f9f91bc799b491146f1","target":"6318ed26b8415943b81b6e73282339cc","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.238},{"source":"637df3c30efd5f9f91bc799b491146f1","target":"77064b78bc7458b28cf78b7f3e5eeb73","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.15},{"source":"486baa7031875dadadd3ee7d83a7a94d","target":"175e9a9b50aa586bac9fc5c9eb0aec22","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.071},{"source":"817a3ffd1c285c55b8a2eb938ba85f4e","target":"02f7fb51dd405cf7a36ebcdf961d23fd","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.3},{"source":"c2c0e7d5306a5897b8bc22c8fce99645","target":"142d56d7638b53659e24beadb82517fb","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.4},{"source":"c2c0e7d5306a5897b8bc22c8fce99645","target":"175e9a9b50aa586bac9fc5c9eb0aec22","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.083},{"source":"c2c0e7d5306a5897b8bc22c8fce99645","target":"f11c75d47c2e53a4b94fc0a39810c80e","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.071},{"source":"1106a1b9f3515e9290d700aa9e21f35b","target":"ef96446cbb435bdc87f8701e55712374","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.5},{"source":"5547843881175855b026a2df0af18c7a","target":"9df51276f50f503db60d7f16830fbc6b","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.625},{"source":"b857f1ae5f585f12bed9b48d888ca7ab","target":"175e9a9b50aa586bac9fc5c9eb0aec22","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.1},{"source":"26a42a6a9311514da6bbc37bba8457ec","target":"135ad2f14b9a517fae601d9d928531c3","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.667},{"source":"8d34e1fc3d105707a21910c27e565b82","target":"e9707c46872b54fc8d0ab6898b5a8b66","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.833},{"source":"8d34e1fc3d105707a21910c27e565b82","target":"f633af356dbf5f378271c0dc2800d089","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.571},{"source":"8d34e1fc3d105707a21910c27e565b82","target":"37bd6a5426ca5981bb60f3765730bd51","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.571},{"source":"f11c75d47c2e53a4b94fc0a39810c80e","target":"0d964edb4353516da96e5aee3d389c8c","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.25},{"source":"37bd6a5426ca5981bb60f3765730bd51","target":"f633af356dbf5f378271c0dc2800d089","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.667},{"source":"37bd6a5426ca5981bb60f3765730bd51","target":"e9707c46872b54fc8d0ab6898b5a8b66","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.667},{"source":"8669e2f0da50530491a3851cadce7e1a","target":"0bf8bc0d26195a2da57b02e5fc304610","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.125},{"source":"e9707c46872b54fc8d0ab6898b5a8b66","target":"f633af356dbf5f378271c0dc2800d089","relation":"co-occurs","confidence":"EXTRACTED","confidence_score":0.667}],"companies":[{"id":"lab_0","company":"World Labs","thesis":"A world model is a generative, multimodal, and interactive system that produces geometrically, physically, and dynamically consistent 3D worlds \u2014 the \"spatial intelligence\" layer that language models fundamentally lack (today's LLMs are \"eloquent but inexperienced, knowledgeable but ungrounded\"). Spatial reasoning is \"the scaffolding upon which cognition is built,\" making world modeling an exponentially harder problem than language modeling because it must honor semantic meaning, geometric structure, and physical law simultaneously."},{"id":"lab_1","company":"Google DeepMind","thesis":"A world model is an AI system that understands the world well enough to simulate aspects of it \u2014 predicting how an environment evolves and how actions change it \u2014 while modeling causality rather than replaying clips. DeepMind frames world models as \"a key stepping stone on the path to AGI,\" because they let agents be trained in an unlimited curriculum of rich, interactive simulated environments."},{"id":"lab_2","company":"NVIDIA","thesis":"A world (foundation) model is a neural network that ingests text, image, video, and motion/action data to predict and generate physics-aware video of the future state of an environment, modeling spatial relationships and physical interactions. NVIDIA argues WFMs are \"as fundamental as large language models\" and are the critical infrastructure for Physical AI \u2014 letting robots and autonomous vehicles learn from cheap synthetic data and generalize from simulation to the real world rather than from costly real-world collection."},{"id":"lab_3","company":"Meta AI / FAIR (Fundamental AI Research)","thesis":"A world model is a learned internal representation of how the physical world evolves \u2014 rich enough to understand observations, predict consequences of hypothetical actions (\"an internal simulator\"), and plan sequences of actions toward a goal. Meta argues this must be learned by predicting abstract representations in latent space (JEPA), not by reconstructing pixels or next tokens; LeCun's framing is that autoregressive LLMs are fundamentally insufficient for grounded, human/animal-like intelligence because they don't model the world."},{"id":"lab_4","company":"Wayve","thesis":"A world model learns representations of the environment and its future dynamics, giving an autonomous vehicle a structured understanding of its surroundings to make informed decisions. Wayve argues such generative world models are essential because real-world data collection alone cannot economically cover the rare, safety-critical long-tail scenarios driving systems must handle, so models that can imagine and re-drive controllable futures are the path to safer embodied AI."},{"id":"lab_5","company":"Decart","thesis":"A world model is an interactive video experience generated end-to-end by a transformer on a frame-by-frame basis \u2014 internally simulating physics, rules, and graphics while responding to user input in real time, rather than rendering via a classic game engine. Decart argues fast transformer inference is the missing link that turns affordable, high-quality generative real-time video into a new fundamental interface, ultimately replacing engines for entertainment, gaming, and physical-AI training."},{"id":"lab_6","company":"Runway","thesis":"A world model is an AI system that builds an internal representation of an environment and uses it to simulate future events within it. Runway argues that to generate realistic video at all, a model must implicitly learn physics, motion, and the dynamics of the visual world \u2014 so video generation is itself an early form of world simulation, and general world models are the next frontier extending generative AI beyond media into robotics, science, and industrial simulation."},{"id":"lab_7","company":"Odyssey (odyssey.ml / odysseyml)","thesis":"A world model is a causal, action-conditioned system that predicts the next state of the world one step at a time from the current state, recent history, and the latest user action \u2014 not a fixed video sequence. Odyssey argues this matters because \"language is a thin, biased slice of reality,\" whereas video contains trillions of observations of physics, causality, and human behavior, making it the data source rich enough to train genuinely general, interactive intelligence spanning gaming, film, robotics, science, and beyond."},{"id":"lab_8","company":"Microsoft Research","thesis":"A world model is a generative model that learns the underlying dynamics, physics, and structure of an interactive 3D environment directly from large-scale human gameplay data, so it can generate consistent, diverse, and persistent future sequences from observed states and actions. Microsoft frames this jointly with human behavior \u2014 a \"World and Human Action Model\" \u2014 and argues it matters because such a model lets you \"play inside the model\" and supports creative ideation rather than merely replicating a game."},{"id":"lab_9","company":"Luma AI","thesis":"Luma argues AGI is fundamentally multimodal and \"reality is the dataset of AGI\" \u2014 to simulate the universe, language and text alone are insufficient, so models must be jointly trained over all signal (text, video, audio, images). For them a world model is a multimodal system that understands and simulates complex physical systems, and their generative video models (Dream Machine, Ray) are the practical, commercial expression of that world-model research path."},{"id":"lab_10","company":"1X Technologies","thesis":"A world model is a learned, data-driven simulator: a model trained directly on real robot video+action data that predicts how an environment evolves in response to a robot's actions. 1X argues such a \"neurally-simulated\" twin absorbs the full complexity of the real world without manual asset creation, solving robotics' core evaluation bottleneck \u2014 reproducibly testing humanoid policies across millions of scenarios \u2014 and, more recently, doubling as a generative engine that lets NEO \"imagine\" a task before acting."},{"id":"lab_11","company":"Physical Intelligence (\u03c0)","thesis":"For Physical Intelligence, a \"world model\" is largely implicit and behavioral rather than an explicit predictive simulator: physical understanding is acquired by training a single foundation model on embodied experience across many robots, plus internet-scale vision-language pretraining, so the policy generalizes to new settings the way LLMs generalize over text. They argue the hard problem in robotics is not dexterity but generalization, and that physical intelligence emerges from diverse pretraining rather than hand-built models of physics. Notably, with \u03c00.7 they have begun bolting on an explicit \"lightweight world model\" that generates synthetic visual subgoals at inference time to aid visual generalization."},{"id":"lab_12","company":"OpenAI","thesis":"OpenAI frames a world model as an emergent, learned simulator of the physical world that arises from scaling generative video, not from hand-coded physics or 3D priors. They argue that training large video-generation models on internet-scale visual data is \"a promising path towards building general purpose simulators of the physical world,\" and that such simulation capability is critical for AI that deeply understands reality. World-modeling, in their telling, is a side-effect of scale: predict pixels well enough across spacetime and 3D consistency, object permanence, and physical interaction emerge."},{"id":"lab_13","company":"Niantic Spatial","thesis":"A world model of physical places must be ground truth, not plausible generation: Niantic Spatial argues spatial intelligence is the next frontier of AI, and that a Large Geospatial Model \u2014 built from billions of posed real-world scans \u2014 supplies the verifiable, centimeter-accurate, persistent understanding of actual places that LLMs and generative World Foundation Models lack. In their framing, the LGM \"extrapolates locally by interpolating globally,\" giving machines a shared, georeferenced map of reality to perceive, comprehend, and navigate the physical world."},{"id":"lab_14","company":"Waabi","thesis":"Waabi believes a world model is a high-fidelity, neural closed-loop simulator that digitally clones the real world from raw sensor logs and lets the full autonomy stack act inside it with near-zero domain gap \u2014 so that safety can be proven scientifically through simulation rather than accumulated real-world miles. They frame \"Physical AI\" / generative AI in the physical world as requiring a learned simulator that generates the rare, safety-critical scenarios real driving almost never surfaces."},{"id":"lab_15","company":"Tencent (Hunyuan) \u2014 with ByteDance & Kuaishou findings noted","thesis":"Tencent Hunyuan treats a world model as a generative system that produces immersive, explorable, interactive, and causally-grounded environments \u2014 not just pretty video. A world model must simulate game-engine-like dynamics (characters, environments, actions, events) and 3D geometry that stays consistent as you move and act within it, responding controllably to language, keyboard, and mouse input. (Comparators: ByteDance's VideoWorld argues a world model can be learned purely from unlabeled video \u2014 rules, reasoning, planning \u2014 without text or reward, via a Latent Dynamics Model; Kuaishou frames Kling as a video generator that \"simulates the characteristics of the physical world.\")"},{"id":"lab_16","company":"World Labs","thesis":"A world model is a generative model that learns the statistical structure of space and time the way LLMs learn the structure of text \u2014 how light falls on a surface, how a scene looks from an angle no camera captured, how objects respond to force. World Labs argues spatial intelligence (perceiving, reasoning about, and interacting with persistent 3D environments) is AI's next frontier and the necessary scaffolding beyond language."}]}