Action Success Prediction on AndroidWorld (smallest held-out split)
87.1PrecisionVisCritic
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| VisCriticBase Model=Qwen2.5-VL base, Modality=Visual2026.06 | 87.1 | 83.4 | 85.2 | |
| BacktrackAgentBase Model=Qwen2.5-VL base, Modality=Text2026.06 | 75.6 | 64.3 | 69.5 | |
| GUI-PRABase Model=Qwen2.5-VL base, Modality=Text2026.06 | 73.2 | 67.8 | 70.4 |