Principia benchmark: no video model exceeds 0.42 on physics despite VBench ~0.8

anand_bhattad · x · 2026-09-07

The Principia benchmark evaluates Newtonian physics in video models via calibration-independent relational consistency between paired objects, covering eight phenomena (gravity, friction, pendulum, etc.). Across thousands of generations from six state-of-the-art video generators, none exceeded 0.42 despite 0.8 VBench scores; the best VLM detected physics violations with only 67% accuracy.

Related event: Principia Benchmark Exposes Video Models' Poor Physics Understanding(3 posts)→

Original post →

More from Multimodal

Multimodal channel →