Click-in-target on OOD Evaluation Suite (ScreenSpot-Pro, OSWorld-G, ScreenSpot-V2) (unweighted mean)
60Click-in-Target Accuracy (SS-Pro)Qwen3.5-4B (base)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Qwen3.5-4B (base)Base model=Qwen3.5-4B, Cold-start regime=None, RL regime=None2026.06 | 60 | 56.3 | — | 70 | |
| SFD-4B +RL EarlyBase model=Qwen3.5-2B, Cold-start regime=SFD-4B, RL regime=Early2026.06 | 55.7 | 54.5 | 88.8 | 66.3 | |
| SFD-EMA +RL EarlyBase model=Qwen3.5-2B, Cold-start regime=SFD-EMA, RL regime=Early2026.06 | 55.1 | 51.6 | 89 | 65.2 | |
| SFD-4BBase model=Qwen3.5-2B, Cold-start regime=SFD-4B, RL regime=None2026.06 | 54.7 | 50.8 | 89.5 | 65 | |
| SFD-EMABase model=Qwen3.5-2B, Cold-start regime=SFD-EMA, RL regime=None2026.06 | 53.5 | 51.2 | 88.8 | 64.5 | |
| SFD-EMA +RL LateBase model=Qwen3.5-2B, Cold-start regime=SFD-EMA, RL regime=Late2026.06 | 53.2 | 51 | 85.4 | 63.2 | |
| SFD-4B +RL LateBase model=Qwen3.5-2B, Cold-start regime=SFD-4B, RL regime=Late2026.06 | 53 | 51.4 | 88.9 | 64.4 | |
| SFT +RL EarlyBase model=Qwen3.5-2B, Cold-start regime=SFT, RL regime=Early2026.06 | 52.7 | 52.4 | 87.7 | 64.3 | |
| Qwen3.5-2B (base)Base model=Qwen3.5-2B, Cold-start regime=None, RL regime=None2026.06 | 52.2 | 47.5 | 83 | 60.9 | |
| SFT +RL LateBase model=Qwen3.5-2B, Cold-start regime=SFT, RL regime=Late2026.06 | 50 | 51.6 | 87.6 | 63.1 | |
| SFTBase model=Qwen3.5-2B, Cold-start regime=SFT, RL regime=None2026.06 | 47.6 | 49 | 85.9 | 60.8 |