Reproduce “TriPhysBench: Evaluating Physical Intelligence for Embodied Agents” 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

Zhongxiang Sun, Chenghao Yang, Haojie Ding, Boheng Zhang, Fan Yang, Qingsong Wen, Kun Wang

TriPhysBench tests embodied agents’ ability to coordinate video measurements, retrieved physical specifications, and calculations across 410 questions from 200 videos. TriPhys trains a tool-using policy on complete solutions through supervised fine-tuning and preference optimization. Gemini-3.6-Flash reaches 0.5038 answer correctness; on 110 held-out-video questions, TriPhys-27B scores 0.5818 versus 0.4941 for its Qwen3.6-27B backbone, a result limited to this test split.

Summary by alphaXiv.

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