New benchmark reveals AI models fail at basic physics, scoring 97.7% on video reconstruction

Researchers introduced VisPhyWorld, a new framework that evaluates AI's physical reasoning by forcing models to generate executable simulator code from videos. Their benchmark, VisPhyBench, contains 209 scenes. While the pipeline itself successfully reconstructs videos 97.7% of the time, experiments show state-of-the-art multimodal LLMs struggle to infer accurate physical parameters and simulate consistent dynamics, revealing a major gap between semantic understanding and true physical reasoning.