Paper 2604.22748
Agentic World Modeling: Foundations, Capabilities, Laws, and Beyond
- Published
- Apr 2026
- Research lab
- Independent
- Citations
- 10
- GitHub
- 295 stars
01 In brief
Summary
This preprint introduces a taxonomy for world models in agentic AI, organized along two axes: capability level (L1 Predictor, L2 Simulator, L3 Evolver) and governing-law regime (physical, digital, social, scientific).
L1 models learn one-step transition operators; L2 composes them into multi-step, action-conditioned rollouts satisfying long-horizon coherence, intervention sensitivity, and constraint consistency; L3 autonomously revises its model from evidence via a design-execute-observe-reflect loop.
The survey synthesizes over 400 works, analyzing methods, failure modes, and evaluation practices across level-regime pairs.
It proposes decision-centric evaluation principles, a minimal reproducible evaluation package (MREP), and architectural guidance.
The authors argue that the future of agentic AI lies in models that internalize governing laws, simulate dynamics, and evolve through active trial-and-error, moving from passive prediction to reshaping environments.
The paper positions itself as a position-driven survey, offering a unifying framework to connect previously isolated research communities in computer vision, reinforcement learning, robotics, NLP, and AI for science.
It also discusses open problems, security concerns, and the potential for meta-world modeling beyond L3, where governing laws themselves become learnable.
The key takeaway is that progress requires not just larger predictors but models that can simulate and continuously evolve, with symbolic representations potentially crucial for L3-level law revision.
The paper includes extensive references and appendices with philosophical motivations, detailed system analyses, and implementation considerations.
The authors are from multiple institutions including HKUST, NUS, Oxford, and NTU, and the work is available as a preprint with associated homepage, repository, and database links.
The paper concludes that the development of…
02 From the paper
Abstract
As AI systems move from generating text to accomplishing goals through sustained interaction, the ability to model environment dynamics becomes a central bottleneck. Agents that manipulate objects, navigate software, coordinate with others, or design experiments require predictive environment models, yet the term world model carries different meanings across research communities. We introduce a "levels x laws" taxonomy organized along two axes. The first defines three capability levels: L1 Predictor, which learns one-step local transition operators; L2 Simulator, which composes them into multi-step, action-conditioned rollouts that respect domain laws; and L3 Evolver, which autonomously revises its own model when predictions fail against new evidence. The second identifies four governing-law regimes: physical, digital, social, and scientific. These regimes determine what constraints a world model must satisfy and where it is most likely to fail. Using this framework, we synthesize over 400 works and summarize more than 100 representative systems spanning model-based reinforcement learning, video generation, web and GUI agents, multi-agent social simulation, and AI-driven scientific discovery. We analyze methods, failure modes, and evaluation practices across level-regime pairs, propose decision-centric evaluation principles and a minimal reproducible evaluation package, and outline architectural guidance, open problems, and governance challenges. The resulting roadmap connects previously isolated communities and charts a path from passive next-step prediction toward world models that can simulate, and ultimately reshape, the environments in which agents operate. Code and resources are available at: https://github.com/matrix-agent/awesome-agentic-world-modeling.