Theory, Benchmarks & Surveys
Theoretical foundations, evaluation benchmarks, and comprehensive survey literature for world models research.
Theory, Explainability & Position Papers
Emergent World Representations
Investigating when and how neural networks spontaneously learn structured, simulation-capable representations of their training environments — from Othello boards to spatial navigation grids.
Li et al., "General agents contain world models"; Gurnee & Tegmark, "Linear Spatial World Models Emerge in LLMs"Causal Reasoning in Transformers
Evidence that next-token prediction yields genuine causal understanding — transformers trained on sequential data develop internal causal world models that support counterfactual reasoning.
Nichani et al., "Transformers Use Causal World Models in Maze-Solving Tasks"Scaling Laws for World Models
Characterizing the compute-optimal strategies for pre-training agents and world models — how model capacity, data scale, and training compute interact to determine downstream performance.
"Scaling Laws for Pre-training Agents and World Models"Video as the Universal Reasoning Substrate
The position that video generation — as the richest single modality — may serve as a universal language for real-world decision making, subsuming planning, prediction, and control.
"Video as the New Language for Real-World Decision Making"Compositional Generative Modeling
The argument that no single monolithic model can capture the full distribution of reality — compositionality at the model level is necessary for robust, generalizable generation.
"Compositional Generative Modeling: A Single Model is Not All You Need"Physics Cognition in Generation
Evaluating whether and how video generation models learn physically plausible dynamics — probing the gap between pixel-level realism and genuine physical understanding.
PhyWorld: "How Far is Video Generation from World Model: A Physical Law Perspective"