List your company

Refusal Induction on Refusal Induction held-in and held-out

55Activation PatchingDifference-in-means
8.220.3532.544.65Feb 17, 2026

Evaluation results

MethodLinks
Difference-in-meansModel=OLMo-13B, alpha=10, k=12, N=1202026.025555415828
Mean steeringModel=Qwen-14B, alpha=10, k=12, N=1202026.02484581820
Mean steeringModel=OLMo-13B, alpha=10, k=12, N=1202026.024849223430
Difference-in-meansModel=Qwen-14B, alpha=10, k=12, N=1202026.024641272025
ReFTModel=Qwen-14B, alpha=10, k=12, N=1202026.0240346235
ReFTModel=OLMo-13B, alpha=10, k=12, N=1202026.022424143127
Difference-in-meansModel=SOLAR-10.7B, alpha=10, k=12, N=1202026.022423201413
Mean steeringModel=SOLAR-10.7B, alpha=10, k=12, N=1202026.0212101266
ReFTModel=SOLAR-10.7B, alpha=10, k=12, N=1202026.021081055