Calibration
ScreenSpot Pro → Video-MME (with subtitle)
Video-MME (with subtitle) is estimated from ScreenSpot Pro with a linear curve fitted on 6 models measured on both: y = 0.4573·x + 0.5277, R² = 0.63, cross-validated error 7.2 pp. It is used for 4 estimates.
| Estimated model | ScreenSpot Pro | Video-MME (with subtitle) | Source |
|---|---|---|---|
| Holo2-235B-A22B | 70.6% | 85.1% | estimated ± 7.2 pp, low confidence |
| Holo2-30B-A3B | 66.1% | 83.0% | estimated ± 7.2 pp, low confidence |
| Holo2-8B | 58.9% | 79.7% | estimated ± 7.2 pp, low confidence |
| Holo2-4B | 57.2% | 78.9% | estimated ± 7.2 pp, low confidence |
Fitted on
| Model | ScreenSpot Pro | Video-MME (with subtitle) |
|---|---|---|
| Claude Opus 4.5 | 45.7% | 77.6% |
| Gemini 3 Pro | 72.7% | 88.4% |
| LFM3-VL-2B | 55.5% | 69.0% |
| Qwen3.5 397B | 65.6% | 87.5% |
| Qwen3.7 Plus | 79.0% | 88.0% |
| Qwen3.8 Max | 84.5% | 90.4% |