Physics benchmark scores 8 video models: top Seedance 2.5 only 57.76/100

jiqizhixin · x · 2026-10-15

Einsia AI (Navers Lab) unveiled a "physics final exam" benchmark for world models: 40 controlled tasks across 9 physical phenomena (collisions, pendulums, optics, buoyancy, phase transitions, EM induction, granular media, surface tension) test whether 8 video generation models actually obey physical laws rather than just produce smooth footage.

Method: GPT Image generates the first frame, models generate the specified physical process, and perception toolkits (CoTracker, SAM2) enable programmatic measurement against theoretical relations like period stability and T-L scaling.

Results are sobering: Seedance 2.5 ranks first with only 57.76/100 on average, followed by MiniMax H3 and Cosmos3. Even visually coherent outputs consistently violate physical laws.

Original post →

More from Multimodal

Multimodal channel →