Original title: Principia: Relational Physics Tests for Video Models
Varun Varma Thozhiyoor, Shivam Tripathi, Venkatesh Babu Radhakrishnan, Anand Bhattad · September 2026
Testing whether AI video generators understand real-world physics is tricky. Measuring exact speeds or sizes usually requires knowing precise camera settings and scales, which AI videos lack. To solve this, researchers created a benchmarkA standardized test used to measure and compare the performance of different models. called Principia. Instead of tracking one object, they compare two objects in the same clip. If both follow the same physical rule, their motions must match in predictable ways, no matter how the camera is set up. Principia tests eight physical phenomena, like gravity and friction, using controlled real-world recordings. The researchers found that today's AI video tools struggle with this. Across thousands of tested videos from six leading generators, none scored higher than 0.42, showing that these models still fail basic relational physics.
You're watching a race between two cyclists on a track, but the video is sped up or slowed down without telling you. If both cyclists follow the same rules of pedaling, their speeds should relate in a predictable way no matter the playback speed. Here, the benchmark checks if the motions of two objects in AI-generated videos relate correctly according to physical laws, even when the video's timing is unknown.
If AI video generators don't understand physics, the videos they create can look wrong in subtle ways, breaking immersion or leading to errors in training data. This benchmark reveals how current models fail at simple Newtonian relationships, pushing developers to improve physical accuracy.
The abstract does not claim that Principia tests all aspects of physical reasoning, only eight specific Newtonian phenomena. It uses videos recorded under controlled protocols, so the results might not apply to all AI-generated videos. Additionally, the approach depends on having paired objects in the scene, which limits its use to scenes with multiple interacting objects.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Principia: Relational Physics Tests for Video Models — Varun Varma Thozhiyoor, Shivam Tripathi, Venkatesh Babu Radhakrishnan et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.