English

Towards Comprehensive Post Safety Alignment of Large Language Models via Safety Patching

Computation and Language 2024-12-18 v2

Abstract

Safety alignment of large language models (LLMs) has been gaining increasing attention. However, current safety-aligned LLMs suffer from the fragile and imbalanced safety mechanisms, which can still be induced to generate unsafe responses, exhibit over-safety by rejecting safe user inputs, and fail to preserve general utility after safety alignment. To this end, we propose a novel post safety alignment (PSA) method to address these inherent and emerging safety challenges, including safety enhancement, over-safety mitigation, and utility preservation. In specific, we introduce \textsc{SafePatching}, a novel framework for comprehensive PSA, where two distinct safety patches are developed on the harmful data to enhance safety and mitigate over-safety concerns, and then seamlessly integrated into the target LLM backbone without compromising its utility. Extensive experiments on four representative aligned LLMs, including LLaMA-2/3, Gemma and Mistral, show that \textsc{SafePatching} achieves a more comprehensive PSA than baseline methods, further optimizing the balance between being helpful and harmless in current aligned LLMs. Also, \textsc{SafePatching} demonstrates its superiority in continual PSA scenarios.

Keywords

Cite

@article{arxiv.2405.13820,
  title  = {Towards Comprehensive Post Safety Alignment of Large Language Models via Safety Patching},
  author = {Weixiang Zhao and Yulin Hu and Zhuojun Li and Yang Deng and Jiahe Guo and Xingyu Sui and Yanyan Zhao and Bing Qin and Tat-Seng Chua and Ting Liu},
  journal= {arXiv preprint arXiv:2405.13820},
  year   = {2024}
}

Comments

25 pages, 10 figures and 14 tables

R2 v1 2026-06-28T16:36:02.041Z