1/ What do video models know about high school physics? Less than you'd think. Led by Varun Varma Thozhiyoor and Shivam Tripathi with Venkatesh Babu, we built Principia: relational physics tests for video models. Eight laws, 500+ real scenes, one simple idea. paper:

@varunvarmat @shivam_tr @rvbabuiisc Congrats, Anand. Looks great! We also looked into this in VDAWorld ( check out HSPBench (

Thanks Ayush for the pointer! HSPBench is quite related to our work. After skimming through it, I see that VDAWorld fills the physics gaps using generative models with physics engines. Our Principia paper provides a formal calibration-free way to benchmark these relational failures directly from generated videos. We will discuss and cite this in our updated paper. For additional context, here’s the first version of our results (#CVPR2026 Findings): Thanks!

2/ How do we verify physics accuracy in generated videos? The main problem is that we cannot determine absolute time simply from the number of generated frames. Next, we often do not know the spatial scale at which the generative models are operating, the camera’s coordinates, or the mass of the objects involved. To address this, our approach shifts away from measuring single objects in isolation, focusing instead on comparing multiple objects to evaluate their relative consistency. For example, two blocks of different masses placed on identical inclines must reach the bottom at the same time. By checking for this relative consistency, we can completely circumvent the need for absolute measurements or other calibration.

3/ For our experiment, we tested six different generators, four vision-language models, several hundred manually recorded and curated scenes, and thousands of hours of compute spent simply generating many videos for testing purposes. None of the models we tested performed well on the Principia benchmark. Summary: Video models are still far from generating physically consistent video. Principia provides an objective evaluation protocol for testing high school physics in generators -- a step towards making physics consistency a go-to evaluation for video models.

4/ Several interesting videos are on our website: paper: PS: Principia is the latest step in a series of work from our group on exploring what generative models actually know (and don't) about the physical world.

@varunvarmat @shivam_tr @rvbabuiisc coooooll

@varunvarmat @shivam_tr @rvbabuiisc The gap says it all: 0.8 on VBench, under 0.42 on Principia. Rendered motion without Newtonian consistency won't get us reliable world models. These generators make things look right, not be right. 💻