Executable World Model

An executable world model is a representation of a physical or agent environment as code that can be executed to advance the simulation state, inspect intermediate computations, verify against observations, and edit individual components. The paradigm contrasts with the dominant neural-world-model approach (world-model, DIT-based-world-model, foundation-world-model) which represents dynamics implicitly in network weights and produces next states via forward passes that cannot be inspected or directly verified.

The executable form offers several advantages: (1) transparency — the simulation logic can be read and audited by humans, which is critical for safety in physical robotics and autonomous systems; (2) gradient-free optimization — code can be improved by search, mutation, or LLM-based edits without backpropagation, making it accessible to non-differentiable environments; (3) editability — individual components of the world can be modified in isolation to test counterfactuals; and (4) verifiable — the simulation can be checked against observations at intermediate steps, not only at the end of a trajectory.

The paradigm has emerged as an active research area in 2026. SceneCode (arXiv 2605.19587) represents indoor scenes as executable programs for editable indoor scenes with articulated objects. VisPhyWorld (arXiv 2602.13294) probes physical reasoning via code-driven video reconstruction. PatchWorld (arXiv 2605.30880) applies gradient-free optimization to executable world models for agent environments. Baba in Wonderland (arXiv 2605.16725) learns executable dynamics online from observations. Executable World Models for ARC-AGI-3 (arXiv 2605.05138) and the follow-up 'Do Coding Agents Need Executable World Models?' (arXiv 2607.15439) examine the role of executable world models in coding agents.

Code-as-World (arXiv 2608.27549) is the most explicit instantiation: it represents physical worlds as executable code and uses an agentic-discovery loop to construct such representations from multimodal observations. The paradigm is adjacent to neural world models (world-model), differentiable physics simulation, and symbolic-physics engines, but distinct in emphasizing code as the representation substrate rather than weights or PDE solvers.