In the paper "Refuse without Refusal," researchers Minji Kim and Hyounghun Kim investigate the phenomenon of false refusals in large language models, where models incorrectly reject benign queries that contain superficially risky language. The authors propose a structural decomposition of safety-tuning datasets into two primary elements: boilerplate refusal statements and explanatory rationales. Through empirical analysis, the study demonstrates that standard refusal statements hinder a model's ability to distinguish between genuine harm and benign intent because they encourage the model to rely on superficial linguistic cues. To mitigate this, the researchers experimented with a "Rationale-Only" training approach, which focuses on training models using only the explanatory component of a safety response rather than the boilerplate refusal. The results indicate that training on rationales alone significantly reduces the rate of false refusals while maintaining a level of safety performance comparable to traditional methods. Furthermore, this Rationale-Only approach proved effective in in-context learning (ICL) configurations and remained compatible with existing inference-time mitigation techniques. The findings suggest that more granular, fine-grained safety supervision datasets—specifically those prioritizing logical reasoning over repetitive refusal templates—are essential for developing aligned agents that effectively balance helpfulness with safety constraints.