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

Researchers introduce DCR (Discernment via Contrastive Refinement), a new method to reduce over-refusal in safety-aligned large language models. The approach helps LLMs better distinguish between genuinely toxic and seemingly toxic prompts, maintaining safety while improving helpfulness without degrading general capabilities.

• →Over-refusal in safety-aligned LLMs causes models to reject benign prompts by misclassifying them as toxic, reducing usability.
• →Previous mitigation strategies create trade-offs where reducing over-refusal typically weakens protection against genuinely harmful content.
• →DCR introduces a contrastive refinement alignment stage that improves LLMs' ability to distinguish truly toxic from superficially toxic prompts.
• →The method effectively reduces over-refusal while preserving safety benefits with minimal impact on general model capabilities.
• →Evaluation across diverse benchmarks demonstrates the approach offers a more principled direction for AI safety alignment.