Calibration
ScreenSpot Pro → MathVision
MathVision is estimated from ScreenSpot Pro with a Michaelis–Menten curve fitted on 7 models measured on both: y = 2.0000·x / (0.95298 + x), R² = 0.51, cross-validated error 12.4 pp. It is used for 4 estimates.
| Estimated model | ScreenSpot Pro | MathVision | Source |
|---|---|---|---|
| Holo2-235B-A22B | 70.6% | 85.1% | estimated ± 12.4 pp, low confidence |
| Holo2-30B-A3B | 66.1% | 81.9% | estimated ± 12.4 pp, low confidence |
| Holo2-8B | 58.9% | 76.4% | estimated ± 12.4 pp, low confidence |
| Holo2-4B | 57.2% | 75.0% | estimated ± 12.4 pp, low confidence |
Fitted on
| Model | ScreenSpot Pro | MathVision |
|---|---|---|
| Claude Opus 4.5 | 45.7% | 74.3% |
| Gemini 3 Pro | 72.7% | 86.6% |
| LFM3-VL-2B | 55.5% | 51.6% |
| Qwen3.5 397B | 65.6% | 88.6% |
| Qwen3.6 Plus | 68.2% | 88.0% |
| Qwen3.7 Plus | 79.0% | 90.3% |
| Qwen3.8 Max | 84.5% | 95.2% |