Research map / Large language models / Retrieval and knowledge
Utilize the Flow Before Stepping into the Same River Twice: Certainty Represented Knowledge Flow for Refusal-Aware Instruction Tuning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CRaFT, a refusal-aware instruction tuning method that adds response certainty to correctness for data selection and uses rehearsal training to track knowledge flow, reducing over-refusal; also introduces the THS metric.
Problem
Refusal-aware instruction tuning (RAIT) with correctness-based labels causes LLMs to over-refuse questions they could answer, due to static and dynamic supervision conflicts.
Approach
Selects RAIT samples using response certainty (negative entropy for MCQA, embedding similarity across sampled answers for OEQA) in addition to correctness, and filters samples using knowledge flow measured by a preliminary rehearsal fine-tuning step (samples whose correctness rises during SFT are dropped from IdK).
Key results
- On MMLU (ID) with LLaMA3-8B-Instruct, CRaFT reaches THS 34.0 versus 20.4 for Cor-RAIT, with error rate 9.6 versus 21.6.
- On TriviaQA (ID) with LLaMA2-7B-Chat, CRaFT improves THS to 23.3 from 19.7 for Cor-RAIT.
- On the OOD NQ set with LLaMA3-8B-Instruct, CRaFT gets THS 0.4 versus -0.1 for Cor-RAIT, while LLaMA2-7B-Chat reaches 6.5 versus 4.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Cor-RAIT with CRaFT, combining certainty and knowledge-flow filtering | LLaMA2-7B-Chat · MMLU (ID) | THS: 11.3 -> 12.5 | Helped |
| replace Cor-RAIT with CRaFT, combining certainty and knowledge-flow filtering | LLaMA2-7B-Chat · ARC-c (OOD) | THS: 11.1 -> 12.6 | Helped |
| replace Cor-RAIT with CRaFT, combining certainty and knowledge-flow filtering | LLaMA2-7B-Chat · TriviaQA (ID) | THS: 19.7 -> 23.3 | Helped |
| replace Init-Refuse with CRaFT, combining certainty and knowledge-flow filtering | LLaMA2-7B-Chat · NQ (OOD) | THS: 04.8 -> 06.5 | Helped |
| replace Cor-RAIT with CRaFT, combining certainty and knowledge-flow filtering | LLaMA3-8B-Instruct · MMLU (ID) | THS: 20.4 -> 34.0 | Helped |
| replace Cor-RAIT with CRaFT, combining certainty and knowledge-flow filtering | LLaMA3-8B-Instruct · ARC-c (OOD) | THS: 12.2 -> 21.4 | Helped |
| replace Cor-RAIT with CRaFT, combining certainty and knowledge-flow filtering | LLaMA3-8B-Instruct · TriviaQA (ID) | THS: 18.8 -> 21.5 | Helped |
| replace Init-Refuse with CRaFT, combining certainty and knowledge-flow filtering | LLaMA3-8B-Instruct · NQ (OOD) | THS: 05.0 -> 00.4 | Hurt |
| add knowledge-flow filtering (ablation) | CRaFT on LLaMA2-7B-Chat · MMLU (ID) | THS: 13.0 -> 12.5 | Hurt |
| add knowledge-flow filtering (ablation) | CRaFT on LLaMA2-7B-Chat · ARC-c (OOD) | THS: 13.5 -> 12.6 | Hurt |
| add knowledge-flow filtering (ablation) | CRaFT on LLaMA3-8B-Instruct · MMLU (ID) | THS: 27.2 -> 34.0 | Helped |
| add knowledge-flow filtering (ablation) | CRaFT on LLaMA3-8B-Instruct · ARC-c (OOD) | THS: 13.9 -> 21.4 | Helped |
Surprising findings
- Van-Tuning (fine-tuning on unmodified data) generally reduces overall capability, suggesting forced-answer instruction tuning undermines the model's inherent refusal ability.
- The OOD OEQA gains of CRaFT over Cor-RAIT are small, and on LLaMA3-8B-Instruct the OOD NQ THS is close to zero.
Limitations
- Evaluated only on MCQA and open-ended QA knowledge tasks; reasoning and multi-turn dialogue not tested.
- Heuristic two-step selection with several thresholds (tau_mu, N_van, N_idk) and rehearsal size; sensitivity is only partly studied.
- OOD improvements for OEQA are limited, and the paper notes performance fluctuates with rehearsal size on MCQA.
Experimental setup
Datasets: MMLU ARC-c TriviaQA Natural Questions
Benchmarks: MMLU ARC-c TriviaQA NQ
Metrics: accuracy (P_c) error rate (P_w) Truthful Helpfulness Score (THS)
Baselines: Init-Basic Init-Refuse Van-Tuning Cor-RAIT R-Tuning
Builds on: R-Tuning Alignment for Honesty Cor-RAIT
Closest papers by meaning
- EVOREFUSE: Evolutionary Prompt Optimization for Evaluation and Mitigation of LLM Over-Refusal to Pseudo-Malicious Instructions (NeurIPS 2025 · similarity 0.88)
- A Closer Look at the Limitations of Instruction Tuning (ICML 2024 · similarity 0.87)
- Latent-space Attacks for Refusal Evasion in Language Models (NeurIPS 2026 · similarity 0.87)
- Just Enough Shifts: Mitigating Over-Refusal in Aligned Language Models with Targeted Representation Fine-Tuning (ICML 2025 · similarity 0.87)
- AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions (NeurIPS 2025 · similarity 0.87)
- Can Language Models Perform Robust Reasoning in Chain-of-thought Prompting with Noisy Rationales? (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.