General agents contain world models

2025/06/02 by Jonathan Richens, Richens, Jonathan, David Abel +5 · 6 voices · 13 citations
#cs.AI #cs.LG #cs.RO #stat.ML

paper · pdf · doi:10.48550/arxiv.2506.01622

Abstract

Are world models a necessary ingredient for flexible, goal-directed behaviour, or is model-free learning sufficient? We provide a formal answer to this question, showing that any agent capable of generalizing to multi-step goal-directed tasks must have learned a predictive model of its environment. We show that this model can be extracted from the agent's policy, and that increasing the agents performance or the complexity of the goals it can achieve requires learning increasingly accurate world models. This has a number of consequences: from developing safe and general agents, to bounding agent capabilities in complex environments, and providing new algorithms for eliciting world models from agents.

Citations

• Language Models Represent Space and Time
• Mastering Diverse Domains through World Models
• A Generalist Agent
• Language Models are Few-Shot Learners
• Mastering Atari, Go, chess and shogi by planning with a learned model
• Machine Theory of Mind
• Theoretical Impediments to Machine Learning With Seven Sparks from the Causal Revolution
• Concrete Problems in AI Safety
• Intelligence without representation

Cited by

• Do World Action Models Generalize Better than VLAs? A Robustness Study
• Qwen-AgentWorld: Language World Models for General Agents
• A Step Toward World Models: A Survey on Robotic Manipulation
• Affordance Representation and Recognition for Autonomous Agents
• World Models Should Prioritize the Unification of Physical and Social Dynamics
• Cogito, Ergo Ludo: An Agent that Learns to Play by Reasoning and Planning
• Internal Planning in Language Models: Characterizing Horizon and Branch Awareness
• TAPO: Transition-Aware Policy Optimization for LLM Agents
• What Capable Agents Must Know: Selection Theorems for Robust Decision-Making under Uncertainty
• LLM-based Agents Suffer from Hallucinations: A Survey of Taxonomy, Methods, and Directions
• CausalARC: Abstract Reasoning with Causal World Models
• Benchmarking LLM-based Agents for Single-cell Omics Analysis
• A "good regulator theorem" for embodied agents

Discussions

• Are world models necessary to achieve human-level agents, or is there a model-free short-cut? Our new #ICML2025 paper tackles this question from first principles, and finds a surprising answer, agents [bsky, 41 points, 2 comments]
• 최근 읍내에서 본 재밌는 논문 인공지능 에이전트가 복잡한 목표를 달성하기 위해서는 반드시 세계 모델(world model)을 학습해야 한다는 것을 수학적으로 증명한 연구 arxiv.org/abs/2506.01622 [bsky, 15 points, 1 comments]
• This week's #PaperILike is "General agents contain world models" (Richens et al., ICML 2025). Brought to my attention by @dabelcs.bsky.social 's beautiful philosophy-laden talk at ICAPS. The explicit [bsky, 8 points, 0 comments]
• Nice paper arxiv.org/abs/2506.01622 [bsky, 5 points, 1 comments]
• General agents contain world models か arxiv.org/abs/2506.01622 [bsky, 0 points, 0 comments]
• arxiv.org/abs/2506.01622 // 꽤 재미있는 이야기 인데 난 이미 연구자료 뽑을때 비슷한 식으로 연구설계를 llm과 한 뒤에 개별요소에 대한 디테일을 정의해 주고 그걸 기반으로 연구방법론을 지정하고 최종적으로 수집한 자료를 교차검증 시킨 뒤에 다시 최종 자료를 기반으로 정리하게 하는 편. [bsky, 0 points, 1 comments]

Related

• Procedural Generalization by Planning with Self-Supervised World Models
• From model-based learning to model-free behaviour with Meta-Interpretive Learning
• From Word to World: Can Large Language Models be Implicit Text-based World Models?
• Language-conditioned world model improves policy generalization by reading environmental descriptions
• Reward-Free Curricula for Training Robust World Models