Agentic World Modeling: Foundations, Capabilities, Laws, and Beyond

The paper introduces a "levels x laws" taxonomy for world modeling in AI, categorizing capabilities into three levels: L1 Predictor, L2 Simulator, and L3 Evolver, each with increasing complexity in modeling environment dynamics. It also identifies four governing-law regimes—physical, digital, social, and scientific—that influence the constraints on world models. This framework synthesizes over 400 works and proposes new evaluation principles and architectural guidance, addressing challenges in developing AI agents capable of dynamic interaction and environment manipulation, which is vital for practitioners aiming to enhance the robustness and adaptability of LLMs and AI systems.