Safety Evaluation on Preregister
3.8Misalignment RateGating Inversion
Evaluation Results
| Method | Links | |||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Gating InversionMitigation Strategy=Realignment Methods, Training Free=true2026.06 | 3.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| re-SFTMitigation Strategy=Realignment Methods, Training Free=false2026.06 | 4.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Activation SteerMitigation Strategy=Realignment Methods, Training Free=true2026.06 | 18.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Misaligned LoRAMitigation Strategy=w/o Mitigation2026.06 | 52.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Alignment Gating (Gating Inversion)Backbone=Qwen3-8B, Protocol=Gating Inversion2026.06 | — | — | 3.8 | 6.2 | 5 | 2.9 | 5.8 | 4.7 | 4.2 | 7.9 | 6.2 | 3.8 | 6.2 | 5.7 | 3.3 | |
| Alignment Gating (Gating Inversion)Backbone=Qwen3-14B, Protocol=Gating Inversion2026.06 | — | — | 2.1 | 2.5 | 1.7 | 0.4 | 0.8 | 1.5 | 2.5 | 2.5 | 2.9 | 2.5 | 1.7 | 2.4 | 2.5 |