Reproduce “Code as Worlds: Agentic Discovery of Executable World Representations for Physical Reasoning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Hanyang Wang, Yimo Cai, Weiliang Chen, Jiawei Chi, Haowen Sun, Qiyu Dai, Yi-Hsin Hung, Xingzhuo Guo and 9 more

The authors represent physical worlds as executable programs with composition, evolution, and appearance components, then build and revise them by comparing simulated rollouts with text or video evidence. They use verified worlds to train vision-language models for quantitative physical reasoning. On QuantiPhy-validation, Code-as-World-VL-9B scores 55.4% threshold-based MRA versus 54.8% for Gemini-3.1 Flash, on 159 pairs using supplied priors.

Key insights

  • An executable representation can join semantic scene structure with explicit physical states and dynamics, making hypotheses editable, simulatable, and testable against observations.
  • Iterative feedback helps use a fixed evaluation budget more effectively: at five evaluations, refinement beats Best-of-5 sampling on most reported reconstruction measures.
  • Text-driven and video-driven simulated worlds provide complementary supervision; together they improve quantitative reasoning beyond image-space training alone.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.