Research map / Large language models / Language, safety and interpretability
Unlearning Isn't Deletion: Investigating Reversibility of Machine Unlearning in LLMs
Xiaoyu Xu, Xiang Yue, Yang Liu, Qingqing Ye, Huadi Zheng, Peizhao Hu, Minxin Du, Haibo Hu
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 |
|---|---|---|---|
| relearn briefly on the forget set after single unlearning | GA, GA+KL, RLabel unlearning · Single unlearning, simple task (Yi-6B) | forget accuracy: GA+KL and RLabel recover F.Acc to approximately 77% | Improved |
| relearn on the cumulative forget set after continual unlearning at moderate LR (5e-6, N=100) | GA continual unlearning · Simple task (Yi-6B) | F.Acc: 9.1 -> 80.0 | Improved |
| relearn after aggressive continual unlearning (LR=3e-5, N=100) | GA continual unlearning · Simple task (Yi-6B) | F.Acc: 0.0 -> 2.1 | No clear effect |
| increase learning rate of continual unlearning from 5e-6 to 3e-5 (ablation) | GA continual unlearning · Mean PCA distance on forget set after relearning (Yi-6B) | mean PCA distance: 1.37 -> 104.58 | Hurt |
| add a retain-set loss (GD or KL) to GA or NPO in single unlearning (ablation) | GA / NPO unlearning · Single unlearning, simple task, LR=3e-6 (Yi-6B) | F.Acc after unlearning: GA 65.4; GA+GD 75.1; GA+KL 71.0; NPO 76.2; NPO+KL 76.3 | Mixed |
| use mean PCA distance before relearning as predictor of recoverability | Diagnostic toolkit · GA continual unlearning (Yi-6B) | recovery of F.Acc: 11.52 vs 133.20 mean PCA distance (forget) pre-relearning | Improved |
| relearning on forget set after continual unlearning pushes accuracy above the original | GA, NPO, NPO+KL, GA+GD · Continual unlearning, simple task (Yi-6B) | F.Acc: 78.9 -> 83.8 (NPO+KL, LR=5e-6, N=100) | Improved |
| min-k%-prob MIA AUC as evaluation of unlearning | MIA metric · Continual unlearning GA, LR=3e-5 (Yi-6B) | MIA AUC: 26.1 -> 74.5 after relearning while F.Acc is 0.0 -> 2.1 | Hurt |
| use GA+GD+WAGLE with relearning restricted to 50% retain set | GA+GD+WAGLE · Yi-6B, LR=2e-5, N=50 | F.Acc / R.Acc after relearning: F.Acc 37.8 -> 46.9; R.Acc 55.9 -> 58.3 | No clear effect |
| use GA+GD+WAGLE with relearning on unrelated data | GA+GD+WAGLE · Yi-6B, LR=4e-5, N=50 | F.Acc after relearning: 27.8 -> 31.5 | No clear effect |
| recover forgotten knowledge via prompt attacks, jailbreaking, quantization or in-context recovery instead of relearning (ablation) | Unlearned models · Reversible catastrophic forgetting regime | recovery of forgotten knowledge | Hurt |
| relearn on forget set vs retain or unrelated data (ablation) | Relearning probe · Sample efficiency (Appendix) | recovery speed | Improved |
Closest papers by meaning
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.93)
- Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.93)
- LLM Unlearning via Loss Adjustment with Only Forget Data (ICLR 2025 · similarity 0.92)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.92)
- Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Difference (NeurIPS 2024 · similarity 0.92)
- On Large Language Model Continual Unlearning (ICLR 2025 · similarity 0.92)
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.
