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

TL;DR

This work tackles the over-refusal observed in Refusal-Aware Instruction Tuning by diagnosing static and dynamic conflicts between samples labeled as vanilla or IdK. It introduces CRaFT, a two-stage framework that (i) queries the model’s knowledge state and monitors knowledge flow via rehearsal training, and (ii) constructs refusal-aware training data guided by correctness, certainty, and knowledge flow to reduce conflicts. Empirical results on MCQA and OEQA tasks show CRaFT improves the Truthful Helpfulness Score (THS) over baselines, demonstrating better balance between reliability and helpfulness and improved generalization to OOD settings. The approach provides a practical pathway to more trustworthy LLM refusals with explicit mechanisms to align training signals with the model’s evolving knowledge, and includes code and data release for reproducibility.

Abstract

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 occurs when similar samples within the LLM's feature space receive differing supervision signals (original vs. modified "I don't know"). Dynamic conflict arises as the LLM's evolving knowledge during SFT enables it to answer previously unanswerable questions, but the now-answerable training samples still retain the original "I don't know" supervision signals from the initial LLM state, leading to inconsistencies. 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 Tuning (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. Code and data will be released at https://github.com/opendatalab/CRaFT .

Figures (9)

• Figure 1: Previous RAIT methods resulted in significant over-refusal, while our CRaFT mitigates this issue, enhancing the LLM's reliability and helpfulness.
• Figure 2: Two causes of over-refusal: (a) Static conflict means the similar samples in the LLM's feature space being assigned different labels (original vs. modified "I don't know"). (b) Dynamic conflict arises since the LLM's knowledge state evolves during SFT, turning initally unknown questions to knowns, while the target answer remains IdK. These conflicts cause the trained LLM to misclassify known questions as unknown, resulting in over-refusal.
• Figure 4: RAIT Data Construction of Cor-RAIT, CorCer-RAIT and CRaFT. Cor-RAIT partitions data based on accuracy $\mu$ and a threshold $\tau_{\mu}$. For CorCer-RAIT, $D_{\text{van}}$ is derived from samples with accuracy exceeding the threshold and the highest certainty, while $D_{\text{idk}}$ consists of samples with accuracy below the threshold and the lowest certainty. CRaFT employs a two-stage process: in the first stage, data with $\Delta\mu > 0$ is excluded through knowledge queries; the second stage follows the same procedure as CorCer-RAIT.
• Figure 5: CRSS of Different RAIT Samples.
• Figure 6: The Framework of CRaFT: Stage 1 queries knowledge state and flow, while Stage 2 constructs RAI data and tunes.