AI

World Models' Last Exam in Physics

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/10/2026, 03:18:25 UTC.

Briefings

A new benchmark called World Models' Last Exam in Physics evaluates video generation models by testing physical consistency across 40 controlled tasks involving mechanics, optics, and thermodynamics. Experiments on eight models across 1,280 videos showed persistent physical inconsistencies, with the best model scoring 57.76 out of 100. The benchmark uses interpretable measurements rather than reference videos and achieves higher agreement with human judgments than a vision-language model baseline.

Key points

• The benchmark World Models' Last Exam in Physics uses 40 controlled tasks covering mechanics, optics, fluids, thermal phenomena, electromagnetism, and surface tension.
• Experiments on eight video generation models across 1,280 videos revealed persistent physical inconsistencies and a best model score of 57.76 out of 100.

Context

This measurement-based benchmark addresses concerns that video world models produce physically inconsistent sequences, providing an interpretable basis for diagnosing physical inconsistencies in embodied AI systems.

Limitations

The benchmark relies on controlled tasks and synthetic video evaluations; direct physical tests focus on mechanics rather than broader physical consistency. Results are based on 1,280 videos from 8 models, and the benchmark provides an interpretable basis for tracking progress toward physically consistent video world models.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 24/10/2026, 22:46:21 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeiccmncw5zujvvh74hfdq32x5a32sj4dgwz7njac6dkvjtjscxjkda