Calibration
SimpleVQA → ScreenSpot Pro
ScreenSpot Pro is estimated from SimpleVQA with a offset logistic curve fitted on 5 models measured on both: y = 0.8505 + (0.0000 − 0.8505) / (1 + exp(−28.00·(x − 0.9089))), R² = 0.97, cross-validated error 2.5 pp. It is used for 6 estimates.
| Estimated model | SimpleVQA | ScreenSpot Pro | Source |
|---|---|---|---|
| dots3-note Preview | 72.5% | 84.6% | estimated ± 2.5 pp, medium confidence |
| Grok 4.20 | 57.4% | 85.0% | estimated ± 2.5 pp, low confidence |
| LFM2.5-VL-3B | 35.4% | 85.1% | estimated ± 2.5 pp, low confidence |
| Qwen3.6-27B | 56.1% | 85.0% | estimated ± 2.5 pp, low confidence |
| Qwen3.6-35B-A3B | 58.9% | 85.0% | estimated ± 2.5 pp, low confidence |
| Step 3.7 Flash | 79.2% | 81.9% | estimated ± 2.5 pp, medium confidence |