Evaluating physical reasoning in video generation models has long been a challenge, as absolute motion measurements are often confounded by frame rate, object scale, and camera calibration. A new benchmark called Principia sidesteps these issues by focusing on relational consistency between paired objects that obey the same physical laws.
Principia spans eight phenomena—gravity, restitution, friction, rotational inertia, projectile motion, momentum, pendulum, and mass-spring oscillation—across translational, rotational, collisional, and oscillatory dynamics. It uses real-world scenes recorded under controlled protocols to provide a robust testbed.
The benchmark introduces a calibration-independent consistency score that measures physical violations directly in image space. When applied to six state-of-the-art video generators, none exceeded a score of 0.42, despite all models scoring around 0.8 on the popular VBench benchmark. This stark contrast highlights a significant gap in physical understanding.
Additionally, the study evaluated vision-language models on their ability to detect relational physics violations. The best model achieved only 67% accuracy, with most performing near chance level, indicating that current AI systems struggle with even basic physical reasoning.