Large language models often memorize harmful or private content that is mixed in with useful knowledge. Traditional methods to remove this content require listing everything to delete, which is impractical because the unwanted information may be hidden. This paper uses the model's own mistakes, hallucinations where it makes unsupported claims, to find what should be forgotten.
The authors introduce Hallucination-Guided Unlearning (HGU). They detect hallucinated spans by comparing model outputs to trusted evidence and checking internal consistency. Then they trace those spans to specific neurons in the model and edit only those neurons using small gates, while preserving the model's overall performance.
On the TOFU benchmark, HGU achieved the best trade-off between forgetting and retaining useful knowledge, with average scores of 0.7221 and 0.7699. This matters because it removes overfitted memories precisely without retraining, making models safer while keeping them capable.