Calibration
Video-MME (with subtitle) → RealWorldQA
RealWorldQA is estimated from Video-MME (with subtitle) with a offset logistic curve fitted on 10 models measured on both: y = 0.7813 + (0.9386 − 0.7813) / (1 + exp(−30.48·(x − 0.8912))), R² = 0.73, cross-validated error 2.8 pp. It is used for 0 estimates.
Fitted on
| Model | Video-MME (with subtitle) | RealWorldQA |
|---|---|---|
| Claude Opus 4.5 | 77.6% | 77.0% |
| Gemini 3 Pro | 88.4% | 83.3% |
| GPT-5.2 | 86.0% | 83.3% |
| Kimi K2.5 | 87.4% | 81.0% |
| LFM3-VL-2B | 69.0% | 79.3% |
| Qwen3.5 397B | 87.5% | 83.9% |
| Qwen3.6-27B | 87.7% | 84.1% |
| Qwen3.6-35B-A3B | 86.6% | 85.3% |
| Qwen3.7 Plus | 88.0% | 86.9% |
| Qwen3.8 Max | 90.4% | 88.0% |