HARC: Coupling Harmfulness and Refusal Directions for Robust Safety Alignment
New fine-tuning method secures LLMs by linking harmfulness and refusal vectors for superior attack resistance.
Understanding how aligned LLMs internally represent safety is critical for diagnosing alignment vulnerabilities, as it explains why jailbreaks succeed and informs the design of robust alignment strategies. Prior work shows that aligned LLMs encode harmfulness and refusal as separ...
ArXiv CS.AI
Source
Generated insight by OnAirToday AI
By identifying and coupling the internal representations of harmfulness and refusal, this research provides a mechanistic blueprint for neutralizing jailbreak vulnerabilities without compromising model utility. This approach shifts safety alignment from generic behavioral training to surgical, subspace-specific interventions that remain robust across diverse model architectures.
Comments
Keep comments relevant and respectful. No links, advertising, self-promotion, or spam.