Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning
Shengyuan Hu, Yiwei Fu, Steven Z. Wu, Virginia Smith
Abstract
Machine unlearning is a promising approach to mitigate undesirable memorization of training data in ML models. However, in this work we show that existing approaches for unlearning in LLMs are surprisingly susceptible to a simple set of benign relearning attacks. With access to only a small and potentially loosely related set of data, we find that we can "jog" the memory of unlearned models to reverse the effects of unlearning. For example, we show that relearning on public medical articles can lead an unlearned LLM to output harmful knowledge about bioweapons, and relearning general wiki information about the book series Harry Potter can force the model to output verbatim memorized text. We formalize this unlearning-relearning pipeline, explore the attack across three popular unlearning benchmarks, and discuss future directions and guidelines that result from our study. Our work indicates that current approximate unlearning methods simply suppress the model outputs and fail to robustly forget target knowledge in the LLMs.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
26 top-tier papers have cited it since. Know what they found before your reviewers do.
Lune can draw on
the full paper · the top-tier papers that cite it · related work
Free to start. No credit card required.
Lineage
- Unlearning Isn't Deletion: Investigating Reversibility of Machine Unlearning in LLMs cites The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning
- Unlearning Isn't Deletion: Investigating Reversibility of Machine Unlearning in LLMs cites From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization
- Large Language Model Unlearning cites Extracting Training Data from Large Language Models
- Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignment cites Direct Preference Optimization: Your Language Model is Secretly a Reward Model
- Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning cites Large Language Model Unlearning
- Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning cites Towards Unbounded Machine Unlearning
- Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning cites Direct Preference Optimization: Your Language Model is Secretly a Reward Model
- Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning cites The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning
- The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning cites Large Language Model Unlearning
- The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning cites Towards Unbounded Machine Unlearning
- The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning cites Direct Preference Optimization: Your Language Model is Secretly a Reward Model
- From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization cites Towards Unbounded Machine Unlearning
- From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization cites The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning
Cited by top-tier papers26
- Simplicity Prevails: Rethinking Negative Preference Optimization for LLM UnlearningChongyu Fan, Jiancheng Liu, Licong Lin, Jinghan Jia et al.NeurIPS 2025 · 182 citations
- The Attacker Moves Second: Stronger Adaptive Attacks Bypass Defenses Against LLM Jailbreaks and Prompt InjectionsMilad Nasr, Nicholas Carlini, Chawin Sitawarin, Sander V. Schulhoff et al.USENIX Security 2026 · 134 citations
- Unlearning Isn't Deletion: Investigating Reversibility of Machine Unlearning in LLMsXiaoyu Xu, Xiang Yue, Yang Liu, Qingqing Ye et al.ICML 2026 · 36 citations
- Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignmentCameron Tice, Puria Radmard, Samuel Ratnam, Andy Kim et al.ICML 2026 · 22 citations
- From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space RegularizationShoaib Ahmed Siddiqui, Adrian Weller, David Krueger, Gintare Karolina Dziugaite et al.NeurIPS 2025 · 16 citations
Builds on12
- Direct Preference Optimization: Your Language Model is Secretly a Reward ModelRafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D. Manning et al.NeurIPS 2023 · 10,924 citations
- Extracting Training Data from Large Language ModelsNicholas Carlini, Florian Tramèr, Eric Wallace, Matthew Jagielski et al.USENIX Security 2021 · 2,866 citations
- The WMDP Benchmark: Measuring and Reducing Malicious Use with UnlearningNathaniel Li, Alexander Pan, Anjali Gopal, Summer Yue et al.ICML 2024 · 390 citations
- Large Language Model UnlearningYuanshun Yao, Xiaojun Xu, Yang LiuNeurIPS 2024 · 365 citations
- Towards Unbounded Machine UnlearningMeghdad Kurmanji, Peter Triantafillou, Jamie Hayes, Eleni TriantafillouNeurIPS 2023 · 363 citations
Related papers
- To Each (Textual Sequence) Its Own: Improving Memorized-Data Unlearning in Large Language ModelsGeorge-Octavian Barbulescu, Peter TriantafillouICML 2024 · 41 citations
- SUA: Stealthy Multimodal Large Language Model Unlearning AttackXianren Zhang, Hui Liu, Delvin Ce Zhang, Xianfeng Tang et al.EMNLP 2025
- MUSE: Machine Unlearning Six-Way Evaluation for Language ModelsWeijia Shi, Jaechan Lee, Yangsibo Huang, Sadhika Malladi et al.ICLR 2025
- Towards Effective Evaluations and Comparisons for LLM Unlearning MethodsQizhou Wang, Bo Han, Puning Yang, Jianing Zhu et al.ICLR 2025
- Catastrophic Failure of LLM Unlearning via QuantizationZhiwei Zhang, Fali Wang, Xiaomin Li, Zongyu Wu et al.ICLR 2025
