ResearcharXiv cs.AI

HARC: Coupling Harmfulness and Refusal Directions for Robust Safety Alignment

#safety#alignment#machine learning#AI#research

English

The paper introduces HARC (Harmfulness-And-Refusal Coupling), a fine-tuning method aimed at improving the safety alignment of large language models (LLMs) by coupling harmfulness and refusal directions. The study reveals that jailbreaks exploit separable harmfulness and refusal directions, and HARC demonstrates a strong trade-off between robustness and usability across various model families without degrading general capability.

中文

本文介绍了HARC（有害性与拒绝耦合），这是一种旨在通过耦合有害性和拒绝方向来改善大型语言模型（LLMs）安全对齐的微调方法。研究表明，越狱攻击利用可分离的有害性和拒绝方向，HARC在不同模型家族中展示了强大的鲁棒性和可用性之间的平衡，而不会降低一般能力。