6K samples across 162 physics phenomena, rendered as videos from analytic SciPy simulations and PyBullet scenes covering rigid-body, articulated, fluid, and contact-rich dynamics.
AI models can state physics laws but rarely infer them from video
PhySimCode, a new benchmark, tests ten multimodal LLMs on turning physics videos into runnable Python simulations, revealing a large gap between verbal physics knowledge and video-grounded reasoning.
Big Tech
Souraja Kundu · Aditya Gupta · Joel Julin · Yizhou Zhao · Liuyue Xie · Laszlo A. Jeni
Carnegie Mellon University · Amazon AGI
Research Digest··3 min read
6K procedurally generated physics simulation videos paired with code, parameters, and physics chain-of-thought explanations.
Why this paper
From Amazon AGI and Carnegie Mellon University
In one line
State-of-the-art multimodal LLMs recover only 7.7% of ground-truth physics parameters from video, exposing a large video-to-simulation code gap.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§