PhyRubric Benchmark: Top Text-to-Video Model Seedance 2.5 Clears Just 44% on Physical Commonsense

UIC researchers on November 11 released PhyRubric, a rubric-grounded diagnostic framework evaluating 10 text-to-video models on 245 captions across mechanics, optics, heat, electricity and acoustics. The strongest model, Seedance 2.5, produces videos that are both faithful and physically plausible in only 44.2% of cases. Open-source models never depict the intended event in 20-44.5% of runs versus 5.7-8.6% for proprietary models, and their 8B fine-tuned PhyRubric-Judge matches frontier Gemini 4 Argon on evaluation correlation.