Calibration
VideoMMMU → MathVision
MathVision is estimated from VideoMMMU with a Michaelis–Menten curve fitted on 10 models measured on both: y = 2.0000·x / (1.14920 + x), R² = 0.77, cross-validated error 6.2 pp. It is used for 0 estimates.
Fitted on
| Model | VideoMMMU | MathVision |
|---|---|---|
| Claude Opus 4.5 | 84.4% | 74.3% |
| dots3-note Preview | 86.8% | 87.7% |
| Gemini 3 Pro | 87.6% | 86.6% |
| GPT-5.2 | 85.9% | 83.0% |
| Kimi K2.5 | 86.6% | 84.2% |
| LFM3-VL-2B | 49.6% | 51.6% |
| Qwen3.5 397B | 84.7% | 88.6% |
| Qwen3.6 Plus | 84.0% | 88.0% |
| Qwen3.7 Plus | 85.4% | 90.3% |
| Qwen3.8 Max | 88.7% | 95.2% |