Bernard Ghanem

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection

Rank-One Safety Injection (ROSI), a white-box method that amplifies a model's safety alignment by permanently steering its activations toward the refusal-mediating subspace, is proposed, suggesting that targeted, interpretable weight steering is a cheap and potent mechanism to improve LLM safety, complementing more resource-intensive fine-tuning paradigms.