Grounding-error detection on ScreenSpot-Pro
87.4Recalldual-channel contrastive classification
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| dual-channel contrastive classificationConfiguration=Image + text veto2026.03 | 87.4 | 88.8 | 91.5 | 87.8 | |
| dual-channel contrastive classificationConfiguration=Image cross-app (max)2026.03 | 83.1 | 88.8 | 88.9 | 84.5 | |
| dual-channel contrastive classificationConfiguration=Image per-app (max)2026.03 | 79.7 | 81.9 | 85.9 | 80.3 | |
| dual-channel contrastive classificationConfiguration=Dual agree (both must match)2026.03 | 69.9 | 71.5 | 77.9 | 70.3 |