Refusal steering on N512 objective checks
94.8Target Success RateResidual Paving
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Residual PavingModel=Qwen2.5-1.5B, Objective=2-obj., Rows=22026.05 | 94.8 | 97.6 | 60.2 | 27.1 | |
| Residual PavingModel=Qwen2.5-1.5B, Objective=3-obj., Rows=22026.05 | 93.2 | 97.6 | 12.2 | 80.6 | |
| Residual PavingModel=Qwen2.5-7B, Objective=2-obj., Rows=22026.05 | 79.9 | 99.2 | 52 | 30.6 | |
| Residual PavingModel=Qwen2.5-7B, Objective=3-obj., Rows=22026.05 | 75.1 | 99.2 | 4.1 | 74.5 |