Calibration
Video-MME (with subtitle) → SimpleVQA
SimpleVQA is estimated from Video-MME (with subtitle) with a inverse Michaelis–Menten curve fitted on 10 models measured on both: y = 1.60743·(x − 0.2809) / (0.2809 + 2.0000 − x), R² = 0.54, cross-validated error 10.8 pp. It is used for 0 estimates.
Fitted on
| Model | Video-MME (with subtitle) | SimpleVQA |
|---|---|---|
| Claude Opus 4.5 | 77.6% | 65.7% |
| Gemini 3 Pro | 88.4% | 73.2% |
| GPT-5.2 | 86.0% | 55.8% |
| Kimi K2.5 | 87.4% | 71.2% |
| LFM3-VL-2B | 69.0% | 36.3% |
| Qwen3.5 397B | 87.5% | 67.1% |
| Qwen3.6-27B | 87.7% | 56.1% |
| Qwen3.6-35B-A3B | 86.6% | 58.9% |
| Qwen3.7 Plus | 88.0% | 81.7% |
| Qwen3.8 Max | 90.4% | 75.0% |