Research map / Large language models / Language, safety and interpretability
Hallucination-Guided Unlearning: Using Hallucination Traces to Reveal Overfitted Memories
Jiayu Zhang, Ziqi Zhong, Yuliang Gai, Canran Xiao
Read the paperCheck your idea against 137,217 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 Hallucination-Guided Unlearning (HGU), which uses the model's own hallucinations as signals to identify overfitted or memorized content to forget in LLMs.
Problem
Selective unlearning in LLMs needs a forget set, but harmful or memorized content is latent and entangled and hard to specify explicitly.
Approach
Treats hallucination traces produced by the model as a signal for what to forget, identifying overfitted memories from the model's own failure modes rather than relying solely on explicit supervision.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| trace hallucinated spans to sparse MLP neurons and gate them under a KL utility constraint | LLM unlearning · TOFU forget05 | Avg (MU, FE): 0.6433 -> 0.7221 vs best baseline RapidUn | Improved |
| use HGU under mixed forget/retain prompts | LLM unlearning · TOFU SepS forget05 | H-Avg: 0.3140 -> 0.4382 vs IDK+AP | Improved |
| apply HGU on WMDP auxiliary Economics | Zephyr-7B · WMDP auxiliary Economics | Forget accuracy: 16.8 -> 14.2 vs RapidUn | Improved |
| raise keep-KL weight lambda_KL (ablation) | HGU · TOFU / WMDP | forgetting vs utility drift | Mixed |
| restrict intervention to traced channel set C (ablation) | HGU · hallucination reduction vs utility | significance of hallucination reduction | Improved |
| HGU vs GA+GD on TOFU forget10 average | HGU · TOFU forget10 | FE: 0.9346 -> 0.8915 | Hurt |
Closest papers by meaning
- BALTO: Balanced Token-Level Policy Optimization for Hallucination Mitigation (NeurIPS 2026 · similarity 0.90)
- Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling (NeurIPS 2025 · similarity 0.89)
- Token-Guard: Towards Token-Level Hallucination Control via Self-Checking Decoding (ICLR 2026 · similarity 0.89)
- ReDeEP: Detecting Hallucination in Retrieval-Augmented Generation via Mechanistic Interpretability (ICLR 2025 · similarity 0.89)
- Learning to Decode Against Compositional Hallucination in Video Multimodal Large Language Models (ICML 2026 · similarity 0.89)
- Detecting and Mitigating Hallucination in Large Vision Language Models via Fine-Grained AI Feedback (AAAI 2025 · similarity 0.89)
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.
