Reproduce “GPUPhysBench: Benchmarking Coding Agents for Correct and Efficient GPU Physics Simulation” 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
Yuchen Sun, Jinjin He, Sinan Wang, Bo Zhu
GPUPhysBench evaluates coding agents on 50 physics-simulation tasks across fluids, deformable solids, and granular materials. Agents implement prescribed numerical methods in CUDA, then face hidden-input correctness checks and timing against expert GPU references. In one-attempt results, Opus and GPT pass all tasks, but Opus exceeds 90% of reference speed on 22%. It tests implementation rather than physical-model selection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.