ChronoPhyBench - Next-State Selection (6 Options): leaderboard
ChronoPhyBench (2026) tests whether multimodal models predict what happens next in real physical-interaction videos rather than relying on language priors. On this board the model sees eight frames of a video with its caption and chooses the next frame from six candidates; the score is accuracy (chance 16.7 percent). Higher is better.
Metric: Accuracy (%) on ChronoPhyBench next-state frame selection with 6 candidates: given a real physical-interaction video context (8 uniformly sampled frames) and its caption, choose the frame that comes next; closed models at temperature 0.1, open models at their recommended sampling; chance 16.7; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 25 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 70.5 | #49 |
| 2 | Doubao-Seed-2.0-Pro-260215 | 65 | #152 |
| 3 | Gemini 3 Pro (Preview) | 63 | #113 |
| 4 | Doubao-Seed-1-8-251228 | 53 | #282 |
| 5 | Gemini 2.5 Pro | 46.63 | #266 |
| 6 | GLM-4.1V-9B (Thinking) | 39.5 | #1060 (GLM-4.1V-9B) |
| 7 | Doubao-Seed-1.6-thinking-250715 | 38.5 | #324 |
| 8 | Gemini 2.5 Flash | 37.85 | #453 |
| 9 | GPT-5.4 | 37.5 | #123 |
| 10 | GPT-5.3 | 34.5 | #197 |
| 11 | Qwen 3 VL 32B Instruct | 25.5 | #585 |
| 12 | Qwen 3 VL 8B (Thinking) | 25 | #887 |
| 13 | InternVL3-8B | 23.5 | #1688 |
| 14 | Qwen 3 VL 8B Instruct | 22.5 | #981 |
| 15 | InternVL3-14B | 18.5 | #1182 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=chronophybench-next-state-selection-6-options · How It Works · Data refreshed daily, snapshot 2026-11-13.