Calibration
MathVision → ScreenSpot Pro
ScreenSpot Pro is estimated from MathVision with a linear curve fitted on 6 models measured on both: y = 1.8439·x + -0.9144, R² = 0.91, cross-validated error 4.5 pp. It is used for 12 estimates.
| Estimated model | MathVision | ScreenSpot Pro | Source |
|---|---|---|---|
| Gemma 4 12B | 79.7% | 55.5% | estimated ± 4.5 pp, medium confidence |
| GPT-5.2 | 83.0% | 61.6% | estimated ± 4.5 pp, medium confidence |
| Kimi K2.6 | 87.4% | 69.7% | estimated ± 4.5 pp, medium confidence |
| Kimi K3 | 94.3% | 82.4% | estimated ± 4.5 pp, medium confidence |
| Qwen3.5-122B-A10B | 86.2% | 67.5% | estimated ± 4.5 pp, medium confidence |
| Qwen3.5-27B | 86.0% | 67.1% | estimated ± 4.5 pp, medium confidence |
| Qwen3.5-35B-A3B | 83.9% | 63.3% | estimated ± 4.5 pp, medium confidence |
| Qwen3.8-27B | 90.0% | 74.5% | estimated ± 4.5 pp, medium confidence |
| Qwen3.8-Flash-Next | 90.6% | 75.6% | estimated ± 4.5 pp, medium confidence |
| Qwen3.8-Omni-Flash | 91.8% | 77.8% | estimated ± 4.5 pp, medium confidence |
| Seed 2.1 Pro | 92.6% | 79.3% | estimated ± 4.5 pp, medium confidence |
| Seed 2.1 Turbo | 90.1% | 74.7% | estimated ± 4.5 pp, medium confidence |