List your company

Refusal Steering on Phi/Qwen judged grid

33.3Target Success RateGoR/DIM ablation
0.9569.35317.7526.147May 18, 2026

Evaluation results

MethodLinks
GoR/DIM ablationEvidence block=External-method comparator context, Rows=182026.0533.387.657.332.5
Refusal direction ablationEvidence block=External-method comparator context, Rows=182026.0532.687.656.633.8
Refusal direction ActAddEvidence block=External-method comparator context, Rows=182026.0530.690.162.239.7
GoR/DIM act-subEvidence block=External-method comparator context, Rows=182026.05299062.536.6
GoR-RDO ablationEvidence block=External-method comparator context, Rows=182026.0527.998.574.829.1
RepIt projectionEvidence block=External-method comparator context, Rows=182026.055.398.568.99.8
No interventionEvidence block=External-method comparator context, Rows=182026.052.298.41003.5