Calibration
VideoMMMU → SimpleVQA
SimpleVQA is estimated from VideoMMMU with a Michaelis–Menten curve fitted on 11 models measured on both: y = 2.0000·x / (1.70155 + x), R² = 0.56, cross-validated error 8.7 pp. It is used for 0 estimates.
Fitted on
| Model | VideoMMMU | SimpleVQA |
|---|---|---|
| Claude Opus 4.5 | 84.4% | 65.7% |
| dots3-note Preview | 86.8% | 72.5% |
| Gemini 3 Pro | 87.6% | 73.2% |
| GPT-5.2 | 85.9% | 55.8% |
| Kimi K2.5 | 86.6% | 71.2% |
| LFM3-VL-2B | 49.6% | 36.3% |
| Qwen3.5 397B | 84.7% | 67.1% |
| Qwen3.6-27B | 84.4% | 56.1% |
| Qwen3.6-35B-A3B | 83.7% | 58.9% |
| Qwen3.7 Plus | 85.4% | 81.7% |
| Qwen3.8 Max | 88.7% | 75.0% |