Video World Models Flunk Physics: Best Model Scores 57.76/100 on New 40-Task Benchmark
Mingju Gao · hf · 2026-10-07
Researchers introduce World Models' Last Exam in Physics, a measurement-based benchmark probing physical consistency in video world models—addressing the problem of visually convincing but physically inconsistent generations.
- 40 controlled tasks spanning mechanics, optics, fluids, thermal/phase-change, electromagnetism, and surface tension;
- Each task pairs an initial image and prompt with predefined physical criteria, enabling interpretable tests without reference videos;
- Across 8 video generation models and 1,280 videos, persistent physical inconsistencies emerged, with the best model scoring 57.76/100;
- The evaluator agrees more with human judgments than a direct vision-language baseline, and was validated on synthetic videos with known physical relationships.
The benchmark offers an interpretable, measurement-grounded basis for diagnosing physics failures and tracking progress toward physically consistent world models.
More from Multimodal
- ComfyUI Qwen image enhancer adds ref mask support — Capitan01R- · 2026-10-07
- DRAMA 1.0 launches: edit only the expression layer, swap 10 emotions in existing footage — nikola_mr64990 · 2026-10-07
- video-shotcraft hits 10.5k stars: cinematic product videos via Claude Code + Remotion — tom_doerr · 2026-10-07
- Third-party test: Google's Nano Banana 2.1 beats ChatGPT Images, 5x faster and cheaper — prajdabre · 2026-10-07
- Single Image to Full 3D Scene: Adaptive Chunking Extends Object Generators to Outdoor Rome — Jiraphon Yenphraphai · 2026-10-07
- ByteDance's DuoMatching: few-step video generation wins 80%+ human preference — ByteDance · 2026-10-07