Calibration
ScreenSpot Pro → ERQA
ERQA is estimated from ScreenSpot Pro with a Michaelis–Menten curve fitted on 10 models measured on both: y = 1.3152·x / (0.78626 + x), R² = 0.46, cross-validated error 8.7 pp. It is used for 4 estimates.
| Estimated model | ScreenSpot Pro | ERQA | Source |
|---|---|---|---|
| Holo2-235B-A22B | 70.6% | 62.2% | estimated ± 8.7 pp, low confidence |
| Holo2-30B-A3B | 66.1% | 60.1% | estimated ± 8.7 pp, low confidence |
| Holo2-8B | 58.9% | 56.3% | estimated ± 8.7 pp, low confidence |
| Holo2-4B | 57.2% | 55.4% | estimated ± 8.7 pp, low confidence |
Fitted on
| Model | ScreenSpot Pro | ERQA |
|---|---|---|
| Claude Opus 4.5 | 45.7% | 46.8% |
| Claude Opus 4.6 | 83.1% | 51.6% |
| Gemini 3.1 Pro | 84.4% | 69.4% |
| Gemini 3 Pro | 72.7% | 70.5% |
| GPT-5.4 | 85.4% | 65.4% |
| LFM3-VL-2B | 55.5% | 48.2% |
| Muse Spark | 84.1% | 64.7% |
| Qwen3.5 397B | 65.6% | 67.5% |
| Qwen3.7 Plus | 79.0% | 69.8% |
| Qwen3.8 Max | 84.5% | 77.8% |