Paper ID: 2410.06913

Utilize the Flow before Stepping into the Same River Twice: Certainty Represented Knowledge Flow for Refusal-Aware Instruction Tuning

Runchuan Zhu, Zhipeng Ma, Jiang Wu, Junyuan Gao, Jiaqi Wang, Dahua Lin, Conghui He

Refusal-Aware Instruction Tuning (RAIT) enables Large Language Models (LLMs) to refuse to answer unknown questions. By modifying responses of unknown questions in the training data to refusal responses such as "I don't know", RAIT enhances the reliability of LLMs and reduces their hallucination. Generally, RAIT modifies training samples based on the correctness of the initial LLM's response. However, this crude approach can cause LLMs to excessively refuse answering questions they could have correctly answered, the problem we call over-refusal. In this paper, we explore two primary causes of over-refusal: Static conflict emerges when the RAIT data is constructed solely on correctness criteria, causing similar samples in the LLM's feature space to be assigned different labels (original vs. modified "I don't know"). Dynamic conflict occurs due to the changes of LLM's knowledge state during fine-tuning, which transforms previous unknown questions into knowns, while the training data, which is constructed based on the initial LLM, remains unchanged. These conflicts cause the trained LLM to misclassify known questions as unknown, resulting in over-refusal. To address this issue, we introduce Certainty Represented Knowledge Flow for Refusal-Aware Instructions Construction (CRaFT). CRaFT centers on two main contributions: First, we additionally incorporate response certainty to selectively filter and modify data, reducing static conflicts. Second, we implement preliminary rehearsal training to characterize changes in the LLM's knowledge state, which helps mitigate dynamic conflicts during the fine-tuning process. We conducted extensive experiments on open-ended question answering and multiple-choice question task. Experiment results show that CRaFT can improve LLM's overall performance during the RAIT process. Source code and training data will be released at Github.

Submitted: Oct 9, 2024