AIAny
Icon for item

Code as Worlds: Agentic Discovery of Executable World Representations for Physical Reasoning

Rewrites physical scenes as executable world programs (e.g., MuJoCo scene descriptions) and uses an agentic abductive loop to propose, execute, render, verify, and iteratively refine those programs from videos or text. Verified executable worlds supply scalable physical supervision for training vision–language models.

Introduction

Pixels record appearances, not the underlying ontology that governs dynamics. Code-as-World flips the representation: instead of treating frames as the primary world model, it induces compact, editable, executable world descriptions (scene programs) that can be simulated, rendered, and verified against real observations. That executable abstraction lets you inspect states, parameters, and mechanisms rather than relying on implicit latent predictors.

Key Findings
  • Agentic discovery loop: an LLM-driven agent proposes executable world hypotheses, runs them in a physics engine, renders proxy observations, compares to source video, and iteratively patches programs until they match—turning raw footage into verified, reusable simulatable worlds.
  • Executable supervision: verified programs provide exact physical labels (states, dynamics, parameters) that scale to large datasets and are used to fine-tune vision–language models for quantitative physical reasoning.
  • Empirical gains and release: fine-tuned Code-as-World-VL checkpoints (4B, 9B) and codebase are released; the approach achieves state-of-the-art performance on QuantiPhy validation in the authors' evaluation.
  • Practical pipeline: pairs an interpretable world-evolution module (code) with a video renderer so that knowledge-driven dynamics control long-term consequences while learned video priors supply high-fidelity visuals.
Who it's for and tradeoffs

Great fit if you are a researcher or engineer aiming to (a) extract interpretable physical structure from real videos, (b) generate precise physical supervision for training models, or (c) build editable simulatable environments from footage. Look elsewhere if you need pixel-perfect photorealism without simulation or if you cannot afford a simulation/rendering stack—simulation fidelity and the simulator–real gap limit applicability for high-precision control on real hardware. The method also requires iterated agentic synthesis and verification, which increases engineering complexity compared to purely end-to-end learned predictors.

Method (brief)

The core loop is propose → execute → render → verify → refine. Candidate scene programs encode objects, state variables, and dynamics; a lightweight compiler produces a proxy video which conditions a video model to render observations; verification turns mismatches into counterexamples that guide program patches. Verified worlds become labeled data for supervised training of vision–language models focused on quantitative physical reasoning.

Information

  • Websitearxiv.org
  • AuthorsHanyang Wang, Yimo Cai, Weiliang Chen, Jiawei Chi, Haowen Sun, Qiyu Dai, Yi-Hsin Hung, Xingzhuo Guo, Jinshan Ren, Runmao Yao
  • Published date2026/08/27

More Items

Proposes VLAct, a representation-centric continued pre-training method for Vision-Language-Action models that preserves VLM priors and enforces cross-embodiment action semantics to turn limited robot trajectories into transferable visual-action representations; shows strong gains and sample efficiency on multiple VLA benchmarks using modest compute.

Performs causal, bounded‑memory streaming 3D reconstruction by caching KV features from only the preceding 11 frames, predicting a per‑frame point map and adjacent relative pose, and composing these local predictions into a global trajectory; includes a lightweight rotation refiner and composition‑aware loss to limit drift.

Turns a flow-matching image generator's self-exploration into dense, per-step supervision without a pretrained teacher; it branches the student's next-state into stochastic SDE candidates, scores them against a deterministic self-reference, and applies an advantage-weighted pull–push velocity regression with reward-level fusion for multi-objective alignment.