Safety Evaluation on strongREJECT (Acceptance rate (%))
99.4Acceptance RateMisaligned LoRA
Evaluation Results
| Method | Links | |
|---|---|---|
| Misaligned LoRAMitigation Strategy=w/o Mitigation2026.06 | 99.4 | |
| Activation SteerMitigation Strategy=Realignment Methods, Training Free=true2026.06 | 73.5 | |
| Gating InversionMitigation Strategy=Realignment Methods, Training Free=true2026.06 | 4.2 | |
| re-SFTMitigation Strategy=Realignment Methods, Training Free=false2026.06 | 4 |