Click-in-target on WinDOM in-distribution 2,000-example subsample (held-out)
82.1Click-in-target AccuracySFT
Evaluation Results
| Method | Links | |
|---|---|---|
| SFTBase model=Qwen3.5-2B, Cold-start regime=SFT, RL regime=None2026.06 | 82.1 | |
| SFT +RL LateBase model=Qwen3.5-2B, Cold-start regime=SFT, RL regime=Late2026.06 | 80.8 | |
| SFD-4B +RL LateBase model=Qwen3.5-2B, Cold-start regime=SFD-4B, RL regime=Late2026.06 | 79.7 | |
| SFD-EMA +RL LateBase model=Qwen3.5-2B, Cold-start regime=SFD-EMA, RL regime=Late2026.06 | 79 | |
| SFT +RL EarlyBase model=Qwen3.5-2B, Cold-start regime=SFT, RL regime=Early2026.06 | 77.2 | |
| SFD-EMABase model=Qwen3.5-2B, Cold-start regime=SFD-EMA, RL regime=None2026.06 | 76.8 | |
| SFD-4BBase model=Qwen3.5-2B, Cold-start regime=SFD-4B, RL regime=None2026.06 | 76.5 | |
| SFD-4B +RL EarlyBase model=Qwen3.5-2B, Cold-start regime=SFD-4B, RL regime=Early2026.06 | 76.5 | |
| SFD-EMA +RL EarlyBase model=Qwen3.5-2B, Cold-start regime=SFD-EMA, RL regime=Early2026.06 | 75.4 | |
| Qwen3.5-4B (base)Base model=Qwen3.5-4B, Cold-start regime=None, RL regime=None2026.06 | 71.2 | |
| Qwen3.5-2B (base)Base model=Qwen3.5-2B, Cold-start regime=None, RL regime=None2026.06 | 70 |