Papers with machine unlearning
Knowledge Control for Responsible Generative AI: Bridging Academia, Industry, and Society (2026.acl-tutorials)
Copied to clipboard
Zheyuan Liu, Yixin Wan, Kai-Wei Chang, Meng Jiang, Jieyu Zhao, Nouha Dziri, Yuning Mao, Jia-Chen Gu, Jindong Gu
| Challenge: | This tutorial introduces the foundations of post-training knowledge control and showcases recent frontier methods. |
| Approach: | This tutorial introduces the foundations of post-training knowledge control and showcases recent frontier methods. |
| Outcome: | This tutorial introduces the foundations of post-training knowledge control and showcases recent frontier methods . key motivations and failure modes, harmful generation and stereotype reinforcement, are addressed . core methods such as machine unlearning, knowledge editing, and inference-time interventions are also included . |
LLMs and Copyright Risks: Benchmarks and Mitigation Approaches (2025.naacl-tutorial)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have revolutionized natural language processing, but their widespread use has raised significant copyright concerns. |
| Approach: | This tutorial will provide an overview of relevant copyright principles and their application to AI and examine specific copyright issues in LLM development and deployment. |
| Outcome: | The course will provide an overview of relevant copyright principles and their application to AI, followed by an examination of specific copyright issues in LLM development and deployment. |
Auditing Language Model Unlearning via Information Decomposition (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing approaches to unlearning in language models do not account for information about forgotten data . despite the apparent success of unlearning, information about the forgotten data remains linearly decodable from internal representations. |
| Approach: | They propose an interpretable framework for auditing unlearning using Partial Information Decomposition . they propose a representation-based risk score that can guide abstention on sensitive inputs . |
| Outcome: | The proposed framework can guide abstention on sensitive inputs at inference time. |
Thesis Proposal: Targeted and Unified Cross-Lingual Unlearning from Multilingual Language Models (2026.acl-srw)
Copied to clipboard
| Challenge: | Large language models trained on corpora scraped from the web can reproduce sensitive and copyright-protected data. |
| Approach: | They propose to extend existing benchmarks to multilingual data by compiling parallel translations of question-answer pairs consisting of real-world facts and synthetic personally identifiable information. |
| Outcome: | The proposed dataset will include translations of question-answer pairs consisting of real-world facts and synthetic personally identifiable information. |
Balancing Forget Quality and Model Utility: A Reverse KL-Divergence Knowledge Distillation Approach for Better Unlearning in LLMs (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing methods for unlearning large language models struggle with forget quality and model utility, leading to over-unlearning or partial unlearning. |
| Approach: | They propose a method that uses reverse KL-divergence based knowledge distillation for unlearning to achieve significant forget quality while maintaining model utility. |
| Outcome: | The proposed method outperforms existing methods in forget quality and model utility with larger unlearning datasets. |
Learning to Refuse: Towards Mitigating Privacy Risks in LLMs (2025.coling-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) exhibit remarkable capabilities in understanding and generating natural languages, but can inadvertently memorize private information, posing significant privacy risks. |
| Approach: | They propose to use a dataset to evaluate machine unlearning methods for protecting personal data in a realistic scenario. |
| Outcome: | The proposed model outperforms baseline methods by 5.65 points and protects target individuals’ personal data while maintaining general capabilities. |
Soft Token Attacks Cannot Reliably Audit Unlearning in Large Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Recent work shows that soft token attacks can extract unlearned information from large language models. |
| Approach: | They show that soft token attacks can extract unlearned information from LLMs . |
| Outcome: | The proposed attacks can extract unlearned information from large language models . |
Soft Prompting for Unlearning in Large Language Models (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing ethical and safety considerations for large language models are important for deployment . however, some ethical concerns have been raised due to the presence of private, sensitive, or harmful information in the training data. |
| Approach: | They propose a framework that learns prompt tokens that are prepended to a query to induce unlearning in LLMs. |
| Outcome: | The proposed method improves the trade-off between utility and forgetting for text classification and question-answering. |
Protecting Privacy in Multimodal Large Language Models with MLLMU-Bench (2025.naacl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) and Multimodal Large Language models (MLLMs) trained on vast web corpora can memorize and disclose individuals’ confidential and private data, raising legal and ethical concerns. |
| Approach: | They propose a benchmark to assess unlearning algorithms from multiple perspectives and provide a baseline for existing generative models. |
| Outcome: | The proposed benchmark consists of 500 fictitious profiles and 153 profiles for public celebrities, evaluated from both multimodal (image+text) and unimodal (text) perspectives. |
Multilingual Amnesia: On the Transferability of Unlearning in Multilingual LLMs (2026.eacl-long)
Copied to clipboard
Alireza Dehghanpour Farashah, Aditi Khandelwal, Marylou Fauchard, Zhuan Shi, Negar Rostamzadeh, Golnoosh Farnadi
| Challenge: | Existing studies on unlearning in multilingual large language models focus on monolingual settings, typically English. |
| Approach: | They propose to use a multilingual data and concept unlearning model to investigate the problem . they extend benchmarks for factual knowledge and stereotypes into ten languages . |
| Outcome: | The proposed model is able to unlearning in 10 languages across five languages and resource levels. |
Unintended Memorization of Sensitive Information in Fine-Tuned Language Models (2026.eacl-long)
Copied to clipboard
Marton Szep, Jorge Marin Ruiz, Georgios Kaissis, Paulina Seidl, Rüdiger von Eisenhart-Rothe, Florian Hinterwimmer, Daniel Rueckert
| Challenge: | Large Language Models (LLMs) on sensitive datasets carry a substantial risk of unintended memorization and leakage of Personally Identifiable Information (PII) prior studies have analyzed memorizing dynamics in LLMs during pre-training and fine-tuning. |
| Approach: | They investigate the vulnerability of PII that appears only in model inputs, not in training targets. |
| Outcome: | The proposed methods show that post-training methods provide more consistent privacy-utility trade-offs . |
JPU: Bridging Jailbreak Defense and Unlearning via On-Policy Path Rectification (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are vulnerable to diverse jailbreak attacks despite extensive safety alignment . |
| Approach: | They propose a method to rectify dynamic jailbreak paths towards safety anchors by dynamically mining on-policy adversarial samples to expose vulnerabilities and identify jailbreak path. |
| Outcome: | The proposed model significantly improves jailbreak resistance against dynamic attacks while maintaining its utility. |
Machine Unlearning of Pre-trained Large Language Models (2024.acl-long)
Copied to clipboard
| Challenge: | Using curated datasets, we establish a robust benchmark for unlearning performance, demonstrating that these methods are over 105 times more computationally efficient than retraining. |
| Approach: | They propose a framework for machine unlearning in pre-trained LLMs and integrate gradient ascent with gradient descent on in-distribution data to achieve robustness. |
| Outcome: | The proposed framework is over 105 times more efficient than retraining on in-distribution data and provides detailed guidelines for efficient hyperparameter tuning in the unlearning process. |
Unlearning as multi-task optimization: A normalized gradient difference approach with an adaptive learning rate (2025.naacl-long)
Copied to clipboard
Xiaomeng Jin, Zhiqi Bu, Bhanukiran Vinzamuri, Anil Ramakrishna, Kai-Wei Chang, Volkan Cevher, Mingyi Hong
| Challenge: | Existing methods to remove unwanted knowledge from large language models are formulated as minimizing memorization through the loss of the model. |
| Approach: | They propose a normalized gradient difference algorithm that optimizes a forgetting objective and an automatic learning rate scheduler that allows for better control over the trade-off between the objectives. |
| Outcome: | The proposed method improves on TOFU and MUSE datasets while exhibiting stable training. |
Cross-Lingual Unlearning of Selective Knowledge in Multilingual Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Pretrained language models memorize large amounts of information, raising significant safety concerns. |
| Approach: | They propose an approach to machine unlearning for multilingual language models that selectively erases information across different languages while maintaining overall performance. |
| Outcome: | The proposed approach is compared with existing unlearning baselines and set a new standard for secure and adaptable multilingual language models. |
KGA: A General Machine Unlearning Framework Based on Knowledge Gap Alignment (2023.acl-long)
Copied to clipboard
| Challenge: | Recent legislation of the "right to be forgotten" has led to the interest in machine unlearning . MU can be used to forget specific training instances as if they have never existed . |
| Approach: | They propose a general unlearning framework called KGA to induce forgetfulness . they propose several unlearning evaluation metrics with pertinence . |
| Outcome: | The proposed framework improves on large-scale datasets and provides insight into unlearning for NLP tasks. |
Protecting Privacy Through Approximating Optimal Parameters for Sequence Unlearning in Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Language models (LMs) demonstrate exceptional capabilities on tasks, but are vulnerable to extraction attacks. |
| Approach: | They propose Privacy Protection via Optimal Parameters (POP) which induces the model to forget about some of its training data. |
| Outcome: | The proposed method outperforms the state-of-the-art in retaining LM performance on 9 classification and 4 dialogue benchmarks. |
Aligned but Blind: Alignment Increases Implicit Bias by Reducing Awareness of Race (2025.acl-long)
Copied to clipboard
| Challenge: | et al., 2012) show value-aligned language models exhibit stereotypes in word association tasks . ignoring racial nuances can perpetuate subtle biases in LMs . |
| Approach: | They propose a bias mitigation strategy that incentivizes representation of racial concepts in early model layers. |
| Outcome: | The proposed approach incentivizes representation of racial concepts in early model layers . it reduces implicit bias by reducing the number of ambiguous inputs, the authors show . |
Unilogit: Robust Machine Unlearning for LLMs Using Uniform-Target Self-Distillation (2025.findings-acl)
Copied to clipboard
| Challenge: | Extensive experiments on public benchmarks and an in-house e-commerce dataset demonstrate Unilogit’s superior performance in balancing forget and retain objectives, outperforming state-of-the-art methods such as NPO and UnDIAL. |
| Approach: | They propose a self-distillation method that dynamically adjusts target logits to achieve a uniform probability for the target token. |
| Outcome: | Extensive experiments on public benchmarks and an in-house e-commerce dataset demonstrate Unilogit’s superior performance in balancing forget and retain objectives. |
From Domains to Instances: Dual-Granularity Data Synthesis for LLM Unlearning (2026.findings-acl)
Copied to clipboard
Xiaoyu Xu, Minxin Du, Zitong LI, Zi Liang, Zhibiao Guo, Zhang Shiyu, Peizhao Hu, Qingqing Ye, Haibo Hu
| Challenge: | Currently, the evaluation of unlearning is limited due to the lack of granularity in the model. |
| Approach: | They propose a framework for synthesizing high-quality forget sets that exploits the target model per se to elicit data that matches its internal knowledge distribution through seed-guided and adversarial prompting. |
| Outcome: | The proposed framework achieves a superior balance of relevance, diversity, and efficiency across benchmarks. |
LDEDE: LRP-Driven Efficient Detection and Editing Framework for LLM Privacy Neurons (2026.findings-acl)
Copied to clipboard
Zhao Zhengyuan, Cao Lifeng, null Sunhaodong, Shi Haotian, Du Xuehui, Liu Aodi, Niu Lanjie, Yang Xiaocheng
| Challenge: | Existing privacy protection methods fail to cover context-dependent sensitive information and are prone to performance degradation. |
| Approach: | They propose a Layer-wise Relevance Propagation-driven framework for efficient privacy neuron detection and editing. |
| Outcome: | The proposed framework achieves 80% higher efficiency than gradient attribution methods while reducing leakage risks of Phone, Email, and medical privacy by 42.7%–73.5% on average and cutting computational time by 60%–90%. |