PAWBench (VLM Prediction) - Calibration: 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. Conditional total-variation distance for vision-language models predicting outcomes as text without video generation, times 100 and from 0 to 100; lower is better.
Metric: Conditional total-variation distance for vision-language models predicting outcomes as text without video generation, times 100, on a 0 to 100 scale (0-100); lower is better; PAWBench measures whether a system reproduces the reference distribution over physically possible outcomes across 50 rollouts on 25 scenarios per physical track, averaged over scenes that pass the readability gate. Source: arxiv.org. Saturation forecast: Not forecast. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5V Turbo | 34.8 | #342 |
| 2 | Kimi K2.6 | 38.2 | #173 |
| 3 | Gemini 3.5 Flash | 39.4 | #98 |
| 4 | Qwen 3.5 Plus | 40.9 | #227 |
| 5 | GPT-5.5 | 42.3 | #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-calibration · How It Works · Data refreshed daily, snapshot 2026-11-12.