Video world models flunk physics: 8 SoTA models top out at 57.76/100 on new benchmark
JaynitMakwana · x · 2026-10-08
Einsia released World Models' Last Exam in Physics, a benchmark of 40 controlled tasks across 9 physical categories (mechanics, optics, fluids, thermal/phase change, electromagnetism, surface tension), scoring generated videos against real physical laws.
- 8 SoTA video world models tested; best score just 57.76/100
- Free fall, the easiest task, averaged only 26.61/100
- takeaway: current models generate plausible scenes that violate physics, limiting use for embodied AI prediction and planning
- Paper, project page, and eval code are open-sourced on GitHub
Related event: New Physics Benchmark Tests Video World Models, Top Score Just 57.76(3 posts)→
More from Multimodal
- Fish Audio's Drama 3 lets you direct voice acting in plain language, shifts emotion mid-line — omarsar0 · 2026-10-08
- Single-Word Midjourney Prompt: 'Threnody' Yields Mournful AI Art (Series 3/7) — tisch_eins · 2026-10-08
- 500+ Opus 5.5 generated videos with copy-paste prompts, effort Max recommended — yihui_indie · 2026-10-08
- Vector's 3DSPA reward signal fixes physics violations in generative video models — VectorInst · 2026-10-08
- Hands-on with Veed AI video: voice cloning, avatar, and confusing credit rules — IgorCarron · 2026-10-08
- One prompt generates a 20-second world-tour flyover video with ElevenLabs creative tools — umesh_ai · 2026-10-08