GUI Action Prediction and Grounding on OmniACT Desktop
92.71Type AccuracyGUICrafter-3B
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| GUICrafter-3BEvaluation Protocol=Fine-Tuning, Training Stage=only Stage12026.06 | 92.71 | 78.81 | 78.81 | |
| GUI-R1-3BEvaluation Protocol=Fine-Tuning2026.06 | 91.86 | 78.37 | 78.31 | |
| GUICrafter-3BEvaluation Protocol=Fine-Tuning, Training Stage=Stage 1 + Stage 22026.06 | 91.36 | 82.88 | 82.88 | |
| OS-Atlas-7BEvaluation Protocol=Fine-Tuning2026.06 | 90.24 | 62.87 | 56.73 | |
| GUICrafter-3BEvaluation Protocol=Fine-Tuning, Training Stage=only Stage22026.06 | 90.17 | 76.78 | 76.78 | |
| Qwen2.5-VL-7B*Evaluation Protocol=Fine-Tuning, SFT dataset=GUI-R1-3K2026.06 | 86.23 | 80.17 | 79.8 | |
| Qwen2.5-VL-7BEvaluation Protocol=Zero Shot2026.06 | 84.74 | 79.89 | 79.66 | |
| GPT-4oEvaluation Protocol=Zero Shot2026.06 | 79.97 | 63.25 | 50.67 | |
| Qwen2.5-VL-3B*Evaluation Protocol=Fine-Tuning, SFT dataset=GUI-R1-3K2026.06 | 77.62 | 62.54 | 63.76 | |
| UI-R1-3BEvaluation Protocol=Fine-Tuning2026.06 | 73.41 | 64.12 | 63.98 | |
| OS-Atlas-4BEvaluation Protocol=Fine-Tuning2026.06 | 63.3 | 42.55 | 26.94 | |
| Qwen2.5-VL-3BEvaluation Protocol=Zero Shot2026.06 | 56.95 | 47.97 | 46.89 |