PAWBench: Video Generators Fail Probabilistic Alignment for World Modeling

RisingSayak · x · 2026-08-28

Hugging Face and collaborators released PAWBench to evaluate 11 video generators on physical world modeling. The study found that while models generate realistic videos, none accurately capture both the range of valid futures and their correct probabilities. Models often overreact to irrelevant cues while underreacting to actual physical changes. Fixes like clearer instructions and diverse sampling helped partially but did not solve the issue end-to-end.

Related event: PAWBench Reveals Video Generators Fail Probabilistic Alignment Tests(5 posts)→

Original post →

More from Multimodal

Multimodal channel →