PhyRubric: Seedance 2.5 Clears 44.2% on Physics Benchmark

Even the best text-to-video model on the market produces videos that are both faithful to the caption and physically plausible in only 44.20% of cases. That is the headline number from PhyRubric, a diagnostic benchmark out of the University of Illinois Chicago that tests ten T2V systems against 245 captions spanning 34 categories across mechanics, optics, heat, electricity and acoustics.

Seedance 2.5 leads the table at 44.20% joint performance, with Veo 3.1 at 36.20%. The strongest open-source entry, HunyuanVideo-1.5-8B, reaches 14.93%. The gap widens on a second axis the authors call avoidance — how often the model just doesn't depict the requested event. Open-source models skip the event in 20.0–44.5% of runs; proprietary models do so in 5.7–8.6%. The paper's own framing: "physical commonsense remains far from solved."

Errors cluster. The authors report that 40.8% of assessable videos score 1 or 2 on physical commonsense (fatal or core violation) while 52.9% score a clean 5 — "physical errors are all-or-nothing," in their words. Missing-consequence violations account for 24.0% of failures; permanence and count (objects teleporting, duplicating, merging) another 20.4%.

Alongside the leaderboard, the team releases PhyRubric-Judge, an 8B evaluator fine-tuned from Qwen3-VL-8B-Instruct on 45.9K item-level verdicts. It matches proprietary Gemini 4 Argon on human correlation (0.61 SA, 0.32 PC) — a notable data point in a run of 53 video-generation stories we've tracked in the past 90 days where open judges have lagged frontier ones.