OmniPhys - Open-Ended Process Quality: leaderboard
OmniPhys (arXiv 2608.25398v1, Table 3): multimodal physics problems from junior high school to university, mostly from Chinese educational materials, for 26 MLLMs; process scores average DeepSeek-V3.2 and GPT-4 judges. S3 share of reference key reasoning steps recovered on open-ended tasks; percent, higher is better.
Metric: Judged key-step recovery (%; open-ended tasks, S3). Source: arxiv.org. Saturation forecast: Around February 2027. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 77.15 | #82 |
| 2 | Gemini 2.5 Flash | 72.2 | #255 |
| 3 | Seed-1.6 | 70.16 | #272 |
| 4 | GPT-5.2 | 69.42 | #115 |
| 5 | GLM-4.6V | 68.49 | #330 |
| 6 | Qwen 3 VL 235B A22B | 67.05 | #251 |
| 7 | Claude Sonnet 4.5 | 61.11 | #148 |
| 8 | O4 Mini | 56.48 | #185 |
| 9 | GPT-5.1 | 52.59 | #145 |
| 10 | Grok 4 | 50.6 | #189 |
| 11 | GPT-4o | 39.5 | #357 |
| 12 | Qwen 2.5 VL 7B | 20.03 | #641 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=omniphys-open-ended-process-quality · How It Works · Data refreshed daily, snapshot 2026-11-05.