Apple-PI Benchmark Evaluates Video Models via Physical Law Constraints
July 25, 2026
Apple-PI uses a three-stage perception, formulation, and deduction protocol with chain-of-frames prompting to test video models against explicit physical laws. Current state-of-the-art video models achieve a maximum score of only 0.473 on this diagnostic benchmark.
HOW THIS AFFECTS YOU
●
researcherYou can use this framework to evaluate how well video models adhere to physical constraints rather than just visual patterns.