Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Machine unlearning can be made irreversible by projecting harmful information out of a model's hidden states.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

An LLM unlearning method that projects hidden states so harmful information is irreversibly removed while useful knowledge is preserved.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Projection-based unlearning is a real idea, but the 'irreversible' claim is load-bearing and unverified in the abstract. the 4 major comments →

arxiv 2508.15449 v1 pith:RC4TR6DW submitted 2025-08-21 cs.LG cs.AI

Reliable Unlearning Harmful Information in LLMs with Metamorphosis Representation Projection

Chengcan Wu , Zeming Wei , Huanran Chen , Yinpeng Dong , Meng Sun This is my paper
classification cs.LG cs.AI
keywords machine unlearninglarge language modelsrepresentation projectionrelearning attackcontinuous unlearninghidden state spacemodel safetyirreversible projection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Large language models can hold harmful knowledge that ordinary unlearning methods only suppress; the information stays in the weights and can be revived by fine-tuning. This paper argues that treating unlearning as an irreversible geometric operation fixes that. Its method, Metamorphosis Representation Projection (MRP), projects hidden-state vectors of a chosen layer onto the complement of a subspace that carries harmful information, so the information is destroyed rather than hidden. The reported result is that models unlearned this way keep their general abilities, survive repeated unlearning requests, and resist relearning attacks. A reader should care because it shifts the goal of unlearning from behavior modification to information removal.

Core claim

The central claim is that removing harmful information from an LLM can be made irreversible if it is done in representation space rather than by updating weights. MRP identifies a subspace in the hidden states of a specific network layer that encodes the to-be-forgotten content, then applies a projective transformation that maps every hidden vector to the orthogonal complement of that subspace. Because a projection is a rank-reducing, idempotent map, the erased component cannot be reconstructed by later training—there is no remaining trace in that layer's activations to amplify. The paper reports that this yields unlearning that survives further fine-tuning on harmful data, supports continuo

What carries the argument

The mechanism is the Metamorphosis Representation Projection (MRP): a linear operator, typically P = I − V V^T, applied to hidden-state vectors h at selected transformer layers, where V spans the harmful subspace. Its defining properties are idempotence (P² = P) and non-invertibility, which make the deletion permanent within the modified representation. The projection's placement matters: applying it at the layer(s) where harmful information is concentrated is intended to erase that information while leaving the orthogonal residual that carries useful knowledge.

Load-bearing premise

Harmful and useful information occupy separate directions in a model's hidden-state space, so deleting the harmful directions leaves all useful knowledge untouched.

What would settle it

Fine-tune an MRP-unlearned model on a small set of examples of the harmful behavior. If the behavior returns at levels comparable to models that merely suppressed it, the erasure was not irreversible. Alternatively, train a linear probe on the projected hidden states from held-out harmful prompts; high accuracy means harmful information is still present in that layer.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • An unlearned model cannot be forced back into harmful behavior by fine-tuning on the forgotten data, because the representation no longer contains the information needed to do so.
  • Multiple unlearning requests can be applied sequentially to the same model without accumulating the drift or instability typical of gradient-ascent-based unlearning.
  • Because the projection is applied to representations rather than weights, unlearning does not require a costly retraining pass over the full model.
  • If the harmful subspace is correctly identified, general capabilities remain intact since only one component of the representation is removed.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the same low-dimensional separability holds for other categories of protected content, MRP-style projections could be composed to erase several topics independently, giving a modular, testable alternative to data-specific retraining.
  • The irreversibility cuts both ways: a mistakenly chosen harmful subspace cannot be repaired by later fine-tuning, so deployment would need a verification step or a reversible checkpoint before the projection is applied.
  • A natural extension is to measure how the required projection layer shifts with model scale and domain; if that shift is consistent, layer selection could be predicted rather than tuned per model.
  • The method's success on safety data suggests a concrete experiment for private-data forgetting: check whether representations of personally identifiable text are linearly separable enough for the same projection to remove them without collateral loss.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes Metamorphosis Representation Projection (MRP), a machine-unlearning method for LLMs that applies a projective transformation to the hidden-state space of selected layers. The authors argue that by projecting out a subspace associated with harmful information, MRP irreversibly eliminates that information without parametric training, thereby enabling continuous unlearning and resistance to relearning attacks. The abstract claims state-of-the-art unlearning effectiveness and preserved natural performance. The supplied full text, however, is severely corrupted and largely unreadable; equations, dataset details, and experimental tables are not available. This report is therefore based primarily on the abstract, the visible method fragments, and the conceptual claims made in the paper.

Significance. If the central claims were established, MRP would be a conceptually interesting alternative to gradient-based unlearning: a non-parametric, layer-wise projection is computationally cheap, does not modify weights, and offers a clear geometric interpretation. The availability of code is a strength. However, the paper's core claims go beyond what a projection can guarantee: irreversibility in the linear-algebra sense is not information-theoretic erasure, and the assumed linear separability of harmful and useful information is not demonstrated. The current manuscript does not provide enough experimental evidence to support the state-of-the-art claim, so the significance is contingent on substantial revision and additional validation.

major comments (4)
  1. [Abstract and Method (projection construction)] The claim that MRP 'completely eradicates informational traces' and 'defends against relearning attacks' rests on the word 'irreversible' used in the linear-algebra sense. A projection is singular and idempotent, but information about the removed component can persist elsewhere: residual connections carry earlier-layer representations forward, later layers can reconstruct directions, and the model weights remain untouched. The manuscript does not define a threat model or prove information-theoretic erasure. At minimum, either weaken the claim to 'removes a linear subspace at a chosen layer' or provide direct evidence, e.g., fine-tuning the modified model and showing that harmful behavior cannot be re-learned.
  2. [Method (subspace identification)] The method's load-bearing assumption is that harmful and useful information are linearly separable in the chosen layer's activation space, and that the harmful subspace can be reliably estimated from a finite set of harmful prompts. This is not justified. If the subspace overlaps with useful directions, utility will degrade; if it is incomplete, harmful behavior will persist. The paper should report ablations over the subspace dimension, layer choice, and a linear-probe analysis showing that residual harmful information is actually reduced after projection.
  3. [Experiments (missing results)] The supplied manuscript contains no readable quantitative results, baselines, error bars, or attack protocols. The abstract's state-of-the-art claim is therefore unverifiable. The authors must include standard benchmarks (e.g., WMDP, TOFU), comparisons to gradient ascent and negative preference optimization, metrics for unlearning effectiveness and retain performance, and a relearning-attack setup with fine-tuning. Additionally, an attempted recovery via linear probing on the modified model's activations would directly test the 'irreversible' claim.
  4. [Continuous unlearning (sequential forget sets)] The abstract states that MRP 'enables effective continuous unlearning,' but the manuscript does not specify how multiple forget sets are handled. Does each new request add a new direction to the projection subspace? Could later projections interfere with earlier ones? Please provide the sequential update rule and evaluate on multiple forget requests to support the continuous-unlearning claim.
minor comments (4)
  1. [Full text readability] The body of the manuscript is heavily garbled in the supplied version; equations, tables, and citations are not decodable. A clean version is essential for review.
  2. [Abstract and terminology] The term 'metamorphosis' appears in the method name but is never defined. Please clarify what 'metamorphosis' refers to in the context of representation projection.
  3. [Related work] The paper should cite and discuss recent work on representation engineering, linear probing of harmful concepts, and other non-parametric unlearning approaches to position the contribution accurately.
  4. [Reproducibility] The abstract mentions a code repository, but no version or commit identifier is provided. Please include the exact code version and environment details.

Circularity Check

0 steps flagged

No circularity found: MRP's projection-based unlearning is an empirical method; abstract-level claims do not reduce to their inputs by definition.

full rationale

The abstract claims that applying projective transformations in the hidden state space eliminates harmful information while preserving useful knowledge, and that this is irreversible. The irreversibility is a mathematical property of any projection matrix (P^2=P and singular), but the paper does not define 'harmful information' as the projected-out subspace; it treats harmfulness as a behavioral property to be measured by benchmarks. The claim that the identified directions carry harmful information is an empirical/architectural assumption (linear separability), not a definitional equivalence. Without access to the derivation equations (the provided full text is largely unreadable due to encoding corruption), there is no quotable step where a fitted parameter is later called a prediction, nor a self-citation chain that forces the conclusion. The linear-separability premise may be fragile, but fragility is a correctness risk, not circularity. Therefore no circularity is established.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

Only the abstract is available; the ledger reflects the assumptions implied by the abstract.

free parameters (1)
  • Projection subspace and target layer = not specified in the abstract
    The method must choose which hidden layer and which directions to project away; these choices are likely tuned or derived from data and determine the harm versus utility trade-off.
axioms (2)
  • domain assumption Harmful information is concentrated in a low-dimensional subspace of hidden states, separable from useful knowledge.
    The central premise of the projection-based erasure; if false, the projection cannot selectively remove harm.
  • standard math Matrix projections are irreversible: information in the nulled subspace is destroyed.
    Standard linear algebra; the claim of irreversibility relies on this.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Reliable Unlearning Harmful Information in LLMs with Metamorphosis Representation Projection." pith.science (2026). https://pith.science/paper/RC4TR6DW

@misc{pith2026250815449,
  author       = {Pith},
  title        = {Pith review of: Reliable Unlearning Harmful Information in LLMs with Metamorphosis Representation Projection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RC4TR6DW}},
  note         = {Machine review of arXiv:2508.15449}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

While Large Language Models (LLMs) have demonstrated impressive performance in various domains and tasks, concerns about their safety are becoming increasingly severe. In particular, since models may store unsafe knowledge internally, machine unlearning has emerged as a representative paradigm to ensure model safety. Existing approaches employ various training techniques, such as gradient ascent and negative preference optimization, in attempts to eliminate the influence of undesired data on target models. However, these methods merely suppress the activation of undesired data through parametric training without completely eradicating its informational traces within the model. This fundamental limitation makes it difficult to achieve effective continuous unlearning, rendering these methods vulnerable to relearning attacks. To overcome these challenges, we propose a Metamorphosis Representation Projection (MRP) approach that pioneers the application of irreversible projection properties to machine unlearning. By implementing projective transformations in the hidden state space of specific network layers, our method effectively eliminates harmful information while preserving useful knowledge. Experimental results demonstrate that our approach enables effective continuous unlearning and successfully defends against relearning attacks, achieving state-of-the-art performance in unlearning effectiveness while preserving natural performance. Our code is available in https://github.com/ChengcanWu/MRP.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. LLM Unlearning for Cyber Defense: A Survey on Methods, Challenges, and Emerging Threats

    cs.LG 2026-06 conditional novelty 4.0

    Most gradient-based LLM unlearning methods achieve behavioral suppression, not true forgetting, and current benchmarks cannot certify that knowledge has been removed.

Reference graph

Works this paper leans on

47 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    Machine unlearning in digitalized healthcare arena a comprehensive exploration, 2024

    Abraham Oshni Alvandi. Machine unlearning in digitalized healthcare arena a comprehensive exploration, 2024

    work page 2024

  2. [2]

    Qwen technical report

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609 , 2023

    Pith/arXiv arXiv 2023

  3. [3]

    Machine unlearning

    Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE symposium on security and privacy (SP) , pages 141--159. IEEE, 2021

    work page 2021

  4. [4]

    California Consumer Privacy Act of 2018 (CCPA)

    California State Legislature . California Consumer Privacy Act of 2018 (CCPA) . https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=201720180AB375, June 2018. AB-375, Signed into law on June 28, 2018

    work page 2018

  5. [5]

    Towards making systems forget with machine unlearning

    Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In 2015 IEEE symposium on security and privacy , pages 463--480. IEEE, 2015

    work page 2015

  6. [6]

    Unlearn what you want to forget: Efficient unlearning for llms

    Jiaao Chen and Diyi Yang. Unlearn what you want to forget: Efficient unlearning for llms. arXiv preprint arXiv:2310.20150 , 2023

    Pith/arXiv arXiv 2023

  7. [7]

    Efficient model updates for approximate unlearning of graph-structured data

    Eli Chien, Chao Pan, and Olgica Milenkovic. Efficient model updates for approximate unlearning of graph-structured data. In The Eleventh International Conference on Learning Representations , 2022

    2022

  8. [8]

    Sparse autoencoders find highly interpretable features in language models

    Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. arXiv preprint arXiv:2309.08600 , 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Who’s harry potter? approximate unlearning for llms

    Ronen Eldan and Mark Russinovich. Who’s harry potter? approximate unlearning for llms. 2023

    work page 2023

  10. [10]

    Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation

    Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. arXiv preprint arXiv:2310.12508 , 2023

    Pith/arXiv arXiv 2023

  11. [11]

    The qr transformation a unitary analogue to the lr transformation—part 1

    John GF Francis. The qr transformation a unitary analogue to the lr transformation—part 1. The Computer Journal , 4(3):265--271, 1961

    work page 1961

  12. [12]

    Erasing concepts from diffusion models

    Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision , pages 2426--2436, 2023

    work page 2023

  13. [13]

    On large language model continual unlearning

    Chongyang Gao, Lixu Wang, Kaize Ding, Chenkai Weng, Xiao Wang, and Qi Zhu. On large language model continual unlearning. arXiv preprint arXiv:2407.10223 , 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Calculating the singular values and pseudo-inverse of a matrix

    Gene Golub and William Kahan. Calculating the singular values and pseudo-inverse of a matrix. Journal of the Society for Industrial and Applied Mathematics, Series B: Numerical Analysis , 2(2):205--224, 1965

    work page 1965

  15. [15]

    Adaptive machine unlearning

    Varun Gupta, Christopher Jung, Seth Neel, Aaron Roth, Saeed Sharifi-Malvajerdi, and Chris Waites. Adaptive machine unlearning. Advances in Neural Information Processing Systems , 34:16319--16330, 2021

    work page 2021

  16. [16]

    Jogging the memory of unlearned models through targeted relearning attacks

    Shengyuan Hu, Yiwei Fu, Steven Wu, and Virginia Smith. Jogging the memory of unlearned models through targeted relearning attacks. In ICML 2024 Workshop on Foundation Models in the Wild , 2024

    work page 2024

  17. [17]

    Knowledge unlearning for mitigating privacy risks in language models

    Joel Jang, Dongkeun Yoon, Sohee Yang, Sungmin Cha, Moontae Lee, Lajanugen Logeswaran, and Minjoon Seo. Knowledge unlearning for mitigating privacy risks in language models. arXiv preprint arXiv:2210.01504 , 2022

    Pith/arXiv arXiv 2022

  18. [18]

    Soul: Unlocking the power of second-order optimization for llm unlearning

    Jinghan Jia, Yihua Zhang, Yimeng Zhang, Jiancheng Liu, Bharat Runwal, James Diffenderfer, Bhavya Kailkhura, and Sijia Liu. Soul: Unlocking the power of second-order optimization for llm unlearning. arXiv preprint arXiv:2404.18239 , 2024

    Pith/arXiv arXiv 2024

  19. [19]

    Machine unlearning models for medical care and health data privacy in healthcare 6.0

    Varun Kumar and Dipanjan Sujit Roy. Machine unlearning models for medical care and health data privacy in healthcare 6.0. In Exploration of Transformative Technologies in Healthcare 6.0 , pages 273--302. IGI Global Scientific Publishing, 2025

    work page 2025

  20. [20]

    Protecting privacy through approximating optimal parameters for sequence unlearning in language models

    Dohyun Lee, Daniel Rim, Minseok Choi, and Jaegul Choo. Protecting privacy through approximating optimal parameters for sequence unlearning in language models. arXiv preprint arXiv:2406.14091 , 2024

    Pith/arXiv arXiv 2024

  21. [21]

    The wmdp benchmark: Measuring and reducing malicious use with unlearning

    Nathaniel Li, Alexander Pan, Anjali Gopal, Summer Yue, Daniel Berrios, Alice Gatti, Justin D Li, Ann-Kathrin Dombrowski, Shashwat Goel, Long Phan, et al. The wmdp benchmark: Measuring and reducing malicious use with unlearning. arXiv preprint arXiv:2403.03218 , 2024

    Pith/arXiv arXiv 2024

  22. [22]

    A survey on recommendation unlearning: Fundamentals, taxonomy, evaluation, and open questions

    Yuyuan Li, Xiaohua Feng, Chaochao Chen, and Qiang Yang. A survey on recommendation unlearning: Fundamentals, taxonomy, evaluation, and open questions. arXiv preprint arXiv:2412.12836 , 2024

    arXiv 2024

  23. [23]

    Blockchain-enabled trustworthy federated unlearning

    Yijing Lin, Zhipeng Gao, Hongyang Du, Jinke Ren, Zhiqiang Xie, and Dusit Niyato. Blockchain-enabled trustworthy federated unlearning. arXiv preprint arXiv:2401.15917 , 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Rethinking machine unlearning for large language models

    Sijia Liu, Yuanshun Yao, Jinghan Jia, Stephen Casper, Nathalie Baracaldo, Peter Hase, Yuguang Yao, Chris Yuhao Liu, Xiaojun Xu, Hang Li, et al. Rethinking machine unlearning for large language models. Nature Machine Intelligence , pages 1--14, 2025

    work page 2025

  25. [25]

    Learn to explain: Multimodal reasoning via thought chains for science question answering

    Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In Advances in Neural Information Processing Systems , volume 35, pages 2507--2521, 2022

    work page 2022

  26. [26]

    An adversarial perspective on machine unlearning for ai safety

    Jakub ucki, Boyi Wei, Yangsibo Huang, Peter Henderson, Florian Tram \`e r, and Javier Rando. An adversarial perspective on machine unlearning for ai safety. arXiv preprint arXiv:2409.18025 , 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Eight methods to evaluate robust unlearning in llms

    Aengus Lynch, Phillip Guo, Aidan Ewart, Stephen Casper, and Dylan Hadfield-Menell. Eight methods to evaluate robust unlearning in llms. arXiv preprint arXiv:2402.16835 , 2024

    Pith/arXiv arXiv 2024

  28. [28]

    Tofu: A task of fictitious unlearning for llms

    Pratyush Maini, Zhili Feng, Avi Schwarzschild, Zachary C Lipton, and J Zico Kolter. Tofu: A task of fictitious unlearning for llms. arXiv preprint arXiv:2401.06121 , 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Scalable extraction of training data from aligned, production language models

    Milad Nasr, Javier Rando, Nicholas Carlini, Jonathan Hayase, Matthew Jagielski, A Feder Cooper, Daphne Ippolito, Christopher A Choquette-Choo, Florian Tram \`e r, and Katherine Lee. Scalable extraction of training data from aligned, production language models. In The Thirteenth International Conference on Learning Representations , 2025

    work page 2025

  30. [30]

    Privacy risks of general-purpose language models

    Xudong Pan, Mi Zhang, Shouling Ji, and Min Yang. Privacy risks of general-purpose language models. In 2020 IEEE Symposium on Security and Privacy (SP) , pages 1314--1331. IEEE, 2020

    work page 2020

  31. [31]

    Machine unlearning in digital healthcare: Addressing technical and ethical challenges

    Shahnewaz Karim Sakib and Mengjun Xie. Machine unlearning in digital healthcare: Addressing technical and ethical challenges. In Proceedings of the AAAI Symposium Series , volume 4, pages 319--322, 2024

    work page 2024

  32. [32]

    Muse: Machine unlearning six-way evaluation for language models

    Weijia Shi, Jaechan Lee, Yangsibo Huang, Sadhika Malladi, Jieyu Zhao, Ari Holtzman, Daogao Liu, Luke Zettlemoyer, Noah A Smith, and Chiyuan Zhang. Muse: Machine unlearning six-way evaluation for language models. arXiv preprint arXiv:2407.06460 , 2024

    Pith/arXiv arXiv 2024

  33. [33]

    Ununlearning: Unlearning is not sufficient for content regulation in advanced generative ai

    Ilia Shumailov, Jamie Hayes, Eleni Triantafillou, Guillermo Ortiz-Jimenez, Nicolas Papernot, Matthew Jagielski, Itay Yona, Heidi Howard, and Eugene Bagdasaryan. Ununlearning: Unlearning is not sufficient for content regulation in advanced generative ai. arXiv preprint arXiv:2407.00106 , 2024

    Pith/arXiv arXiv 2024

  34. [34]

    Stanford alpaca: An instruction-following llama model, 2023

    Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Stanford alpaca: An instruction-following llama model, 2023

    2023

  35. [35]

    EU general data protection regulation (GDPR): an implementation and compliance guide

    IT Governance Privacy Team et al. EU general data protection regulation (GDPR): an implementation and compliance guide . Packt Publishing Ltd, 2025

    work page 2025

  36. [36]

    Llama 2: Open foundation and fine-tuned chat models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 , 2023

    Pith/arXiv arXiv 2023

  37. [37]

    Orthogonal subspace learning for language model continual learning

    Xiao Wang, Tianze Chen, Qiming Ge, Han Xia, Rong Bao, Rui Zheng, Qi Zhang, Tao Gui, and Xuanjing Huang. Orthogonal subspace learning for language model continual learning. arXiv preprint arXiv:2310.14152 , 2023

    Pith/arXiv arXiv 2023

  38. [38]

    Certified edge unlearning for graph neural networks

    Kun Wu, Jie Shen, Yue Ning, Ting Wang, and Wendy Hui Wang. Certified edge unlearning for graph neural networks. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages 2606--2617, 2023

    work page 2023

  39. [39]

    Large language model unlearning

    Yuanshun Yao, Xiaojun Xu, and Yang Liu. Large language model unlearning. Advances in Neural Information Processing Systems , 37:105425--105475, 2024

    work page 2024

  40. [40]

    Right to be forgotten in the era of large language models: Implications, challenges, and solutions

    Dawen Zhang, Pamela Finckenberg-Broman, Thong Hoang, Shidong Pan, Zhenchang Xing, Mark Staples, and Xiwei Xu. Right to be forgotten in the era of large language models: Implications, challenges, and solutions. AI and Ethics , pages 1--10, 2024

    work page 2024

  41. [41]

    To be forgotten or to be fair: Unveiling fairness implications of machine unlearning methods

    Dawen Zhang, Shidong Pan, Thong Hoang, Zhenchang Xing, Mark Staples, Xiwei Xu, Lina Yao, Qinghua Lu, and Liming Zhu. To be forgotten or to be fair: Unveiling fairness implications of machine unlearning methods. AI and Ethics , 4(1):83--93, 2024

    work page 2024

  42. [42]

    Negative preference optimization: From catastrophic collapse to effective unlearning

    Ruiqi Zhang, Licong Lin, Yu Bai, and Song Mei. Negative preference optimization: From catastrophic collapse to effective unlearning. arXiv preprint arXiv:2404.05868 , 2024

    Pith/arXiv arXiv 2024

  43. [43]

    Machine unlearning by reversing the continual learning

    Yongjing Zhang, Zhaobo Lu, Feng Zhang, Hao Wang, and Shaojing Li. Machine unlearning by reversing the continual learning. Applied Sciences , 13(16):9341, 2023

    work page 2023

  44. [44]

    Representation engineering: A top-down approach to ai transparency

    Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, et al. Representation engineering: A top-down approach to ai transparency. arXiv preprint arXiv:2310.01405 , 2023

    Pith/arXiv arXiv 2023

  45. [45]

    Universal and transferable adversarial attacks on aligned language models

    Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043 , 2023

    Pith/arXiv arXiv 2023

  46. [46]

    Federated trustchain: Blockchain-enhanced llm training and unlearning

    Xuhan Zuo, Minghao Wang, Tianqing Zhu, Lefeng Zhang, Dayong Ye, Shui Yu, and Wanlei Zhou. Federated trustchain: Blockchain-enhanced llm training and unlearning. arXiv preprint arXiv:2406.04076 , 2024

    Pith/arXiv arXiv 2024

  47. [47]

    Federated learning with blockchain-enhanced machine unlearning: A trustworthy approach

    Xuhan Zuo, Minghao Wang, Tianqing Zhu, Lefeng Zhang, Shui Yu, and Wanlei Zhou. Federated learning with blockchain-enhanced machine unlearning: A trustworthy approach. IEEE Transactions on Services Computing , 2025

    work page 2025

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.