Understanding and Mitigating Over-refusal for Large Language Models via Representation Intervention

Abstract

Large language models (LLMs) demonstrate powerful capabilities across various natural language processing tasks,yet their inherent safety vulnerabilities undermine the reliable application of LLMs in real-world scenarios. To enhance LLM safety, various jailbreak defense methods have been proposed to guard against harmful outputs. However, improvements in model safety often come at the cost of severe over-refusal, failing to strike a good balance between safety and usability. This phenomenon is a critical reliability degradation issue in LLM intelligent systems, failing to strike a good balance between safety defense effectiveness and system usability reliability. In this paper, we first analyze the causes of over-refusal from a representation perspective, revealing that LLMs are unable to effectively distinguish between over-refusal samples and malicious samples. Based on this, we propose to mitigate overrefusal by intervening in the safety representation space of LLMs. Our method incorporates two core strategies: (1) OverlapAware Loss Weighting, which determines the erasure weight for malicious samples by quantifying their similarity to overrefusal samples in the representation space, and (2) ContextAware Augmentation, which supplements the necessary context for rejection decisions by adding harmful prefixes before rejection responses. Experiments demonstrate that our method achieves a better trade-off between mitigating over-refusal and maintaining safety, compared with existing approaches. This paper also aims to encourage researchers to consider the reliability of defending methods against jailbreak attacks from both the perspectives of safety and over-refusal.