Mechanistic Unlearning: Robust Knowledge Unlearning and Editing via Mechanistic Localization
Phillip Guo, Aaquib Syed, Abhay Sheshadri, Aidan Ewart, Gintare Karolina Dziugaite
Abstract
Methods for knowledge editing and unlearning in large language models seek to edit or remove undesirable knowledge or capabilities without compromising general language modeling performance. This work investigates how mechanistic interpretability-which, in part, aims to identify model components (circuits) associated to specific interpretable mechanisms that make up a model capability-can improve the precision and effectiveness of editing and unlearning. We find a stark difference in unlearning and edit robustness when training components localized by different methods. We highlight an important distinction between methods that localize components based primarily on preserving outputs, and those finding high level mechanisms with predictable intermediate states. In particular, localizing edits/unlearning to components associated with the lookup-table mechanism for factual recall 1) leads to more robust edits/unlearning across different input/output formats, and 2) resists attempts to relearn the unwanted information, while also reducing unintended side effects compared to baselines, on both a sports facts dataset and the CounterFact dataset across multiple models. We also find that certain localized edits disrupt the latent knowledge in the model more than any other baselines, making unlearning more robust to various attacks.
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.
10 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
- Improving Alignment and Robustness with Circuit Breakers cites Locating and Editing Factual Associations in GPT
- Towards Automated Circuit Discovery for Mechanistic Interpretability cites Locating and Editing Factual Associations in GPT
- On the creation of narrow AI: hierarchy and nonlocality of neural network skills cites Locating and Editing Factual Associations in GPT
- Deep Ignorance: Filtering Pretraining Data Builds Tamper-Resistant Safeguards into Open-Weight LLMs cites Improving Alignment and Robustness with Circuit Breakers
- Deep Ignorance: Filtering Pretraining Data Builds Tamper-Resistant Safeguards into Open-Weight LLMs cites On the creation of narrow AI: hierarchy and nonlocality of neural network skills
- Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuning cites Locating and Editing Factual Associations in GPT
- A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO and Toxicity cites Locating and Editing Factual Associations in GPT
Cited by top-tier papers10
- Deep Ignorance: Filtering Pretraining Data Builds Tamper-Resistant Safeguards into Open-Weight LLMsKyle O'Brien, Stephen Casper, Quentin Anthony, Tomek Korbak et al.ICLR 2026 · 59 citations
- Steering Out-of-Distribution Generalization with Concept Ablation Fine-TuningHelena Casademunt, Caden Juang, Adam Karvonen, Samuel Marks et al.ICML 2026 · 32 citations
- Attention Smoothing Is All You Need For UnlearningSaleh Zare Zade, Xiangyu Zhou, Sijia Liu, Dongxiao ZhuICLR 2026 · 7 citations
- On the creation of narrow AI: hierarchy and nonlocality of neural network skillsEric J. Michaud, Asher Parker-Sartori, Max TegmarkNeurIPS 2025 · 6 citations
- Randomized Antipodal Search Done Right for Data Pareto Improvement of LLM UnlearningZiwen Liu, Huawei Lin, Yide Ran, Denghui Zhang et al.ICLR 2026 · 2 citations
Builds on14
- Locating and Editing Factual Associations in GPTKevin Meng, David Bau, Alex Andonian, Yonatan BelinkovNeurIPS 2022 · 3,415 citations
- Towards Automated Circuit Discovery for Mechanistic InterpretabilityArthur Conmy, Augustine N. Mavor-Parker, Aengus Lynch, Stefan Heimersheim et al.NeurIPS 2023 · 861 citations
- Improving Alignment and Robustness with Circuit BreakersAndy Zou, Long Phan, Justin Wang, Derek Duenas et al.NeurIPS 2024 · 362 citations
- A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO and ToxicityAndrew Lee, Xiaoyan Bai, Itamar Pres, Martin Wattenberg et al.ICML 2024 · 177 citations
- Can Sensitive Information Be Deleted From LLMs? Objectives for Defending Against Extraction AttacksVaidehi Patil, Peter Hase, Mohit BansalICLR 2024 · 167 citations
Related papers
- CLUE: Conflict-guided Localization for LLM Unlearning FrameworkHang Chen, Jiaying Zhu, Xinyu Yang, Wenya WangICLR 2026 · 8 citations
- Large Scale Knowledge WashingYu Wang, Ruihan Wu, Zexue He, Xiusi Chen et al.ICLR 2025
- DKME: Rethinking Coupled Knowledge Memory for Lifelong Model Editing of Large Language ModelsGuanyu Zheng, Zhenyu Wang, Tingting He, Xv Wang et al.ACL Findings 2026
- Is This the Subspace You Are Looking for? An Interpretability Illusion for Subspace Activation PatchingAleksandar Makelov, Georg Lange, Atticus Geiger, Neel NandaICLR 2024 · 48 citations
- REACT: Representation Extraction And Controllable Tuning to Overcome Overfitting in LLM Knowledge EditingHaitian Zhong, Yuhuan Liu, Ziyang Xu, Guofan Liu et al.EMNLP 2025
