Calibration
Video-MME (w/o subtitle) → MLVU (M-Avg)
MLVU (M-Avg) is estimated from Video-MME (w/o subtitle) with a Hill curve fitted on 7 models measured on both: y = 0.0000 + (0.8819 − 0.0000)·x^6.00 / (0.49341^6.00 + x^6.00), R² = 0.89, cross-validated error 4.4 pp. It is used for 0 estimates.
Fitted on
| Model | Video-MME (w/o subtitle) | MLVU (M-Avg) |
|---|---|---|
| Claude Opus 4.5 | 81.4% | 81.7% |
| Gemini 3 Pro | 87.7% | 83.0% |
| GPT-5.2 | 85.8% | 85.6% |
| Kimi K2.5 | 83.2% | 85.0% |
| LFM3-VL-2B | 62.8% | 71.3% |
| Qwen3.5 397B | 83.7% | 86.7% |
| Qwen3.6-35B-A3B | 82.5% | 86.2% |