List your company

Selective Refusal Editing on HarmBench (Gemma-3-4B-IT held-out split)

88.6Edit Refusal RateBase model (no intervention)
-3.33620.53244.468.268May 18, 2026

Evaluation results

MethodLinks
Base model (no intervention)Type=reference2026.0588.610010081.60
Residual Paving thresholded-soft S4/T2Type=learned route2026.05495.587.365.3-16.3
Residual Paving oracle S4/T2Type=diagnostic2026.050.210010081.60