Calibration
Video-MME (with subtitle) → ScreenSpot Pro
ScreenSpot Pro is estimated from Video-MME (with subtitle) with a offset logistic curve fitted on 6 models measured on both: y = 0.5054 + (0.8456 − 0.5054) / (1 + exp(−135.54·(x − 0.8748))), R² = 0.90, cross-validated error 8.4 pp. It is used for 1 estimate.
| Estimated model | Video-MME (with subtitle) | ScreenSpot Pro | Source |
|---|---|---|---|
| MiniMax M3 | 85.4% | 52.5% | estimated ± 8.4 pp, low confidence |
Fitted on
| Model | Video-MME (with subtitle) | ScreenSpot Pro |
|---|---|---|
| Claude Opus 4.5 | 77.6% | 45.7% |
| Gemini 3 Pro | 88.4% | 72.7% |
| LFM3-VL-2B | 69.0% | 55.5% |
| Qwen3.5 397B | 87.5% | 65.6% |
| Qwen3.7 Plus | 88.0% | 79.0% |
| Qwen3.8 Max | 90.4% | 84.5% |