PAWBench (VLM Prediction) - Coverage: leaderboard
PAWBench (2026) evaluates world models on probabilistic alignment: whether repeated generations recover the reference distribution over physically possible outcomes across 25 scenarios per track. Valid-support recovery for vision-language models predicting outcomes as text, in percent; higher is better.
Metric: Coverage (%; PAWBench measures whether a system reproduces the reference distribution over physically possible outcomes across 50 rollouts on 25 scenarios per physical track from repeated VLM text predictions of the outcome, no video generation; valid-support recovery: share of reference-supported outcomes the model produces (%, higher is better; averaged over passing scenes)). Source: arxiv.org. Saturation forecast: Around June 2028. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.5 Flash | 46.6 | #98 |
| 2 | Kimi K2.6 | 45.1 | #173 |
| 3 | GLM-5V Turbo | 39.9 | #342 |
| 4 | Qwen 3.5 Plus | 35.9 | #227 |
| 5 | GPT-5.5 | 34.3 | #48 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=pawbench-vlm-prediction-coverage · How It Works · Data refreshed daily, snapshot 2026-11-12.