PhyGround
A criteria-grounded benchmark for physical reasoning in generative world models.
Physical Reasoning Benchmark for Generative World Models · Northeastern University · Jan 2026 – May 2026
Paper · Code · Website · Dataset · PhyJudge-9B
- Proposed and built PhyGround, a criteria-grounded benchmark for physical reasoning in generative video world models, spanning a taxonomy of 13 physical laws (solid-body mechanics, fluid dynamics, optics) and evaluating 8 text/image-to-video models over 250 physics-aware prompts.
- Led a large-scale human annotation study with 459 annotators, 5,796 complete annotations, and 37.4K+ fine-grained labels, reaching split-half ranking correlation of Spearman’s \(\rho > 0.90\).
- Trained and released PhyJudge-9B, an open physics-specialized VLM judge that achieves 3.3% aggregate relative bias, compared to 16.6% for Gemini-3.1-Pro.