Discern Truth from Falsehood: Reducing Over-Refusal via Contrastive Refinement

TL;DR

This paper introduces DCR, a contrastive refinement method that improves large language models' ability to distinguish truly toxic prompts from superficially toxic ones, reducing over-refusal without sacrificing safety or general capabilities.

Contribution

The paper proposes a novel alignment stage, DCR, that enhances toxicity discernment in LLMs through contrastive learning, addressing over-refusal issues more effectively than prior methods.

Findings

DCR significantly reduces over-refusal in LLMs.

DCR maintains safety and helpfulness of models.

Empirical results show improved toxicity discrimination.

Abstract

Large language models (LLMs) aligned for safety often suffer from over-refusal, the tendency to reject seemingly toxic or benign prompts by misclassifying them as toxic. This behavior undermines models' helpfulness and restricts usability in sensitive or nuanced contexts. While prior work has proposed mitigation strategies such as data augmentation and activation steering, these approaches often face a trade-off: reducing over-refusal typically degrades the model's ability to reject genuinely harmful content. We argue that this issue arises from the ambiguous influence of toxic and seemingly toxic prompts on the model's learning dynamics. To address it, we introduce a preceding alignment stage, DCR: Discernment via Contrastive Refinement. Both theoretically and empirically, we demonstrate that contrastive refinement improves an LLM's capacity to distinguish truly toxic prompts from superficially toxic ones. Evaluation across diverse benchmarks shows that our method effectively reduces over-refusal while preserving the safety benefits of alignment. Importantly, it achieves this with minimal degradation of general capabilities, offering a more principled and robust direction for safety alignment.