EVENT DOSSIER
The paper was published on arXiv and Hugging Face Papers on September 7, 2026.
On September 7, 2026, Minji Kim and Hyounghun Kim published a paper titled “Refuse without Refusal” on arXiv and Hugging Face Papers, proposing a new method for safe fine-tuning of language models. The study indicates that the “standard refusal statements” used in traditional safe fine-tuning lead to excessive reliance on surface features (such as keyword matching), resulting in false refusals for queries that seem harmful but are actually harmless. The research breaks down the response structure into two parts: “standard refusal statements” and “reasons for refusal”. It was found that training solely based on “reasons for refusal” can significantly reduce false refusals while maintaining adequate security performance. Experiments show that this method is also effective in ICL configurations and is compatible with existing mitigation methods for inference. The study emphasizes that constructing a dedicated dataset containing fine-grained safety supervision data is crucial for developing an alignment agent that better balances usefulness and safety.
政府 · 科研机构across 1 days
The paper was published on arXiv and Hugging Face Papers on September 7, 2026.
Minji Kim and Hyounghun Kim published a paper on arXiv and Hugging Face Papers, proposing to break secure fine-tuning responses into two parts: “standard refusal statements” and “reasons for refusal”. The study found that training based only on reasons can reduce false refusals and maintain security performance.
2 reportsMinji Kim 和 Hyounghun Kim 发表的研究指出，安全微调回复中的标准拒绝陈述是导致语言模型产生虚假拒绝的主要驱动因素。该研究将回复分解为“标准拒绝陈述”和“解释理由”，发现前者使模型过度依赖“射击”等表面线索而非查询含义进行判断。仅基于理由进行训练可减少针对看似有害但无害查询的虚假拒绝，同时保持相当的安全性和通用能力，且该优势在上下文学习中持续有效，并与现有推理时缓解方法兼容。
本文提出将安全微调数据中的回复分解为“标准拒绝陈述”与“拒绝理由”两部分，并发现仅基于理由进行训练可减少语言模型在区分有害与良性查询时的误拒现象，同时保持相当的安全性能。实验表明，标准拒绝陈述会诱导模型依赖表面特征，阻碍准确判别；而仅使用理由训练的效果在指令微调（ICL）配置中同样有效，且兼容现有推理时缓解方法。研究强调需要精心构建细粒度安全监督数据集，以开发更好地平衡有用性与安全的对齐代理。