Research map / Large language models / Instruction and fine-tuning
Fine-Tuned LLMs Know They Don’t Know: A Parameter-Efficient Approach to Recovering Honesty
Zeyu Shi, Ziming Wang, Tianyu Chen, Shiqi Gao, Haoyi Zhou, Qingyun Sun, Jianxin Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Hessian-guided compensation (Stage 2) (ablation) | HCNR · avg F1, FT on HotpotQA | F1: 65.96 -> 72.84 | Improved |
| add Hessian-guided compensation (Stage 2) (ablation) | HCNR · avg F1, FT on MedMCQA | F1: 65.68 -> 73.02 | Improved |
| add Hessian-guided compensation (Stage 2) (ablation) | HCNR · avg RF delta, FT on HotpotQA | Refusal delta: +33.09 -> +42.64 | Improved |
| use honesty-priority neuron selection instead of random (Stage 1) (ablation) | HCNR · avg F1, FT on HotpotQA | F1: 65.44 -> 72.84 | Improved |
| use honesty-priority neuron selection instead of random (Stage 1) (ablation) | HCNR · avg F1, FT on MedMCQA | F1: 67.39 -> 73.02 | Improved |
| add task-importance term to neuron priority (ablation) | HCNR · domain accuracy, FT on HotpotQA | domain accuracy: 28.30 -> 30.30 | Improved |
| add task-importance term to neuron priority (ablation) | HCNR · avg F1, FT on HotpotQA | F1: 70.43 -> 72.84 | Improved |
| add task-importance term to neuron priority (ablation) | HCNR · domain accuracy, FT on MedMCQA | domain accuracy: 9.32 -> 10.33 | Improved |
| use both random selection and no compensation (ablation) | HCNR · avg F1, FT on HotpotQA | F1: 72.84 -> 54.21 | Hurt |
| restore honesty with HCNR instead of RAIT | Llama-3.1-8B-Instruct FT on HotpotQA · domain accuracy | accuracy: 27.05 -> 30.30 | Improved |
| use HCNR instead of DPO for recovery time | Llama-3.1-8B-Instruct FT on HotpotQA · time | minutes: 42.78 min -> 3.93 min | Improved |
| use HCNR (256 samples) instead of RAIT (5000 samples) | Llama-3.1-8B-Instruct FT on HotpotQA · avg F1 | F1: 70.58 -> 72.84 | Improved |
Closest papers by meaning
- Enhancing Multiple Dimensions of Trustworthiness in LLMs via Sparse Activation Control (NeurIPS 2024 · similarity 0.88)
- Do LLMs Signal When They’re Right? Evidence from Neuron Agreement (ICML 2026 · similarity 0.88)
- Remarkable Robustness of LLMs: Stages of Inference? (NeurIPS 2025 · similarity 0.88)
- HonestLLM: Toward an Honest and Helpful Large Language Model (NeurIPS 2024 · similarity 0.87)
- The Sparsity Whisperer (NeurIPS 2026 · similarity 0.86)
- Cross-Modal Unlearning via Influential Neuron Path Editing in Multimodal Large Language Models (AAAI 2026 · 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.
