Decoy Direction Optimization

Researchers propose Decoy Direction Optimization, a post-hoc defense that counters Refusal Feature Ablation (RFA), an attack that projects out a linear refusal direction from a language model's residu…