mlsanityData to 14 Oct 2026

Robust LLM Unlearning with MUDMAN: Meta-Unlearning with Disruption Masking And Normalization

arXiv 2506.12484 ↗First posted 14 Jun 2025cs.LG cs.AI cs.CLcited by 4

On the map: LLM safety and security → Machine unlearning

Abstract (arXiv)

Language models can retain dangerous knowledge and skills even after extensive safety fine-tuning, posing both misuse and misalignment risks. Recent studies show that even specialized unlearning methods can be easily reversed. To address this, we systematically evaluate many existing and novel components of unlearning methods and identify ones crucial for irreversible unlearning. We introduce Disruption Masking, a technique in which we only allow updating weights, where the signs of the unlearning gradient and the retaining gradient are the same. This ensures all updates are non-disruptive. Additionally, we identify the need for normalizing the unlearning gradients, and also confirm the usefulness of meta-learning. We combine these insights into MUDMAN (Meta-Unlearning with Disruption Masking and Normalization) and validate its effectiveness at preventing the recovery of dangerous capabilities. MUDMAN outperforms the prior TAR method by 40%, setting a new state-of-the-art for robust unlearning.

Key facts. Cited by 4 papers in the map (data through 14 Oct 2026). Placed in Machine unlearning, one of 12 directions in LLM safety and security. Builds on 26 of its 42 references that are in the map. First posted on arXiv 14 Jun 2025.

Builds on

26 of 42 references.

Closest ancestors · 4

Same direction, newest first

  1. 2024Do Unlearning Methods Remove Information from Language Model Weights?11 Oct
  2. 2024An Adversarial Perspective on Machine Unlearning for AI Safety26 Sep
  3. 2024Large Scale Knowledge Washing26 May
  4. 2024Eight Methods to Evaluate Robust Unlearning in LLMs26 Feb

Other references · 22

  1. 2020Measuring Massive Multitask Language Understandingcited by 7,963
  2. 2017Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networkscited by 4,897
  3. 2023Universal and Transferable Adversarial Attacks on Aligned Language Modelscited by 3,104
  4. 2016Pointer Sentinel Mixture Modelscited by 3,096
  5. 2019Optuna: A Next-generation Hyperparameter Optimization Frameworkcited by 2,501
  6. 2020The Pile: An 800GB Dataset of Diverse Text for Language Modelingcited by 2,411
  7. 2023Pythia: A Suite for Analyzing Large Language Models Across Training and Scalingcited by 1,786
  8. 2023Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!cited by 1,154
  9. 2018Efficient Lifelong Learning with A-GEMcited by 1,058
  10. 2023BeaverTails: Towards Improved Safety Alignment of LLM via a Human-Preference Datasetcited by 860
  11. 2019CodeSearchNet Challenge: Evaluating the State of Semantic Code Searchcited by 654
  12. 2018Progress & Compress: A scalable framework for continual learningcited by 520
10 more
  1. 2024The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearningcited by 509
  2. 2024Alignment faking in large language modelscited by 360
  3. 2024Improving Alignment and Robustness with Circuit Breakerscited by 322
  4. 2024Deliberative Alignment: Reasoning Enables Safer Language Modelscited by 308
  5. 2024A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO and Toxicitycited by 185
  6. 2024Tamper-Resistant Safeguards for Open-Weight LLMscited by 138
  7. 2023DEPN: Detecting and Editing Privacy Neurons in Pretrained Language Modelscited by 79
  8. 2022Self-Destructing Models: Increasing the Costs of Harmful Dual Uses of Foundation Modelscited by 57
  9. 2024Whispering Experts: Neural Interventions for Toxicity Mitigation in Language Modelscited by 32
  10. 2024Large Language Models Relearn Removed Conceptscited by 27

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 2; 2026: 2.

4 papers.

  1. 2025Collapse of Irrelevant Representations (CIR) Ensures Robust and Non-Disruptive LLM Unlearningcited by 3
  2. 2025LLM Unlearning Under the Microscope: A Full-Stack View on Methods and Metricscited by 2
  3. 2026Modeling LLM Unlearning as an Asymmetric Two-Task Learning Problemcited by 1
  4. 2026RepSelect: Robust LLM Unlearning via Representation Selectivitycited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Downgrade to Upgrade: Optimizer Simplification Enhances Robustness in LLM Unlearning
  2. 2025OpenUnlearning: Accelerating LLM Unlearning via Unified Benchmarking of Methods and Metrics
  3. 2026Unlearning in LLMs: Methods, Evaluation, and Open Challenges
  4. 2025Invariance Makes LLM Unlearning Resilient Even to Unanticipated Downstream Fine-Tuning
  5. 2025Unlearning That Lasts: Utility-Preserving, Robust, and Almost Irreversible Forgetting in LLMs
  6. 2025BLUR: A Benchmark for LLM Unlearning Robust to Forget-Retain Overlap
  7. 2025LLM Unlearning Should Be Form-Independent
  8. 2024Towards Effective Evaluations and Comparisons for LLM Unlearning Methods