Calibration
ScreenSpot Pro → MLVU (M-Avg)
MLVU (M-Avg) is estimated from ScreenSpot Pro with a linear curve fitted on 6 models measured on both: y = 0.3231·x + 0.6178, R² = 0.48, cross-validated error 7.6 pp. It is used for 5 estimates.
| Estimated model | ScreenSpot Pro | MLVU (M-Avg) | Source |
|---|---|---|---|
| GPT-6 Astra | 92.7% | 91.7% | estimated ± 7.6 pp, low confidence |
| Holo2-235B-A22B | 70.6% | 84.6% | estimated ± 7.6 pp, low confidence |
| Holo2-30B-A3B | 66.1% | 83.1% | estimated ± 7.6 pp, low confidence |
| Holo2-8B | 58.9% | 80.8% | estimated ± 7.6 pp, low confidence |
| Holo2-4B | 57.2% | 80.3% | estimated ± 7.6 pp, low confidence |
Fitted on
| Model | ScreenSpot Pro | MLVU (M-Avg) |
|---|---|---|
| Claude Opus 4.5 | 45.7% | 81.7% |
| Gemini 3 Pro | 72.7% | 83.0% |
| LFM3-VL-2B | 55.5% | 71.3% |
| Qwen3.5 397B | 65.6% | 86.7% |
| Qwen3.7 Plus | 79.0% | 87.4% |
| Qwen3.8 Max | 84.5% | 90.8% |