Safety Alignment Should Made Be More Than Just A Few Attention Head

ACL ARR 2025 May Submission6143 Authors

Abstract:

Current implementations of safety alignment for LLMs exhibit notable vulnerabilities.
Our investigation shows that these safety mechanisms predominantly depend on a limited subset of attention heads: removing or ablating these heads can severely compromise model safety.
To identify and evaluate these safety-critical components, we introduce RDSHA, a targeted ablation method that leverages the model's refusal direction to pinpoint attention heads mostly responsible for safety behaviors. Further analysis shows that existing jailbreak attacks exploit this concentration by selectively bypassing or manipulating these critical attention heads.
To address this issue, we propose AHD, a novel training strategy designed to promote the distributed encoding of safety-related behaviors across numerous attention heads.
Experimental results demonstrate that AHD successfully distributes safety-related capabilities across a larger number of attention heads. Moreover, evaluations under several mainstream jailbreak attacks show that models trained with AHD exhibit considerably stronger safety robustness, while maintaining overall functional utility.

Paper Type: Long

Research Area: Language Modeling

Research Area Keywords: safety and alignment

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 6143