List your company

Safety Evaluation on Preregister

3.8Misalignment RateGating Inversion
1.85614.97828.141.222Jun 8, 2026

Evaluation results

MethodLinks
Gating InversionMitigation Strategy=Realignment Methods, Training Free=true2026.063.8——————————————
re-SFTMitigation Strategy=Realignment Methods, Training Free=false2026.064.2——————————————
Activation SteerMitigation Strategy=Realignment Methods, Training Free=true2026.0618.2——————————————
Misaligned LoRAMitigation Strategy=w/o Mitigation2026.0652.4——————————————
Alignment Gating (Gating Inversion)Backbone=Qwen3-8B, Protocol=Gating Inversion2026.06——3.86.252.95.84.74.27.96.23.86.25.73.3
Alignment Gating (Gating Inversion)Backbone=Qwen3-14B, Protocol=Gating Inversion2026.06——2.12.51.70.40.81.52.52.52.92.51.72.42.5