Reproduce “PhysCodeGym: Benchmarking Coding Agents for Physically Verifiable 4D Reconstruction” 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

Shuai Shao, Jinchen He, Cunxin Fan, Zihan Guo, Shijian Wang, Qihan Ren, Kangning Zhang, Zunnan Xu and 5 more

PhysCodeGym evaluates whether agents reconstruct videos as runnable Blender scenes whose motion arises from modeled physical processes, rather than scripted replay. Its 156 tasks span rigid-body, liquid, deformable, and wind-driven motion, with event annotations and selected altered-condition tests. GPT-6-Astra leads the reported aggregate at 0.663. The benchmark tests specified, Blender-expressible requirements, not recovery of a unique true scene.

Summary by alphaXiv.

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