Pith. sign in

REVIEW 3 major objections 8 minor 1 cited by

Editing as Unlearning: Are Knowledge Editing Methods Strong Baselines for Large Language Model Unlearning?

T0 review · 3 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper argues that LLM unlearning is a special case of knowledge editing—replace the target answer with a refusal—and shows that WISE and AlphaEdit are strong unlearning baselines, especially for pretrained knowledge.

desk verdict Useful bridge paper, but the 'unlearning' it benchmarks is behavioral refusal, not knowledge erasure, so the strong-baseline claim needs that caveat. read the letter →

arxiv 2505.19855 v1 pith:LF6ZWSGN submitted 2025-05-26 cs.LG

classification cs.LG
keywords LLMunlearningknowledgeeditingasrefusalanswersWISEAlphaEdithumanvaluealignmentquerymerging
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Unlearning, the paper argues, is not a separate task from knowledge editing: it is the special case where the edited answer is a refusal, the 'empty set' of model knowledge. The authors rewrite the editing objective so that forget queries map to a refusal answer ($\emptyset$) while everything else is preserved, which converts any editing method into an unlearning method. On TOFU's factual set and PISTOL, they find that WISE and AlphaEdit—two lifelong editing methods—are strong unlearning baselines, especially for pretrained knowledge, and that their refusal outputs are rated more human-aligned than those of gradient-based unlearning. Two recipes, self-improvement and query merging, adapt editing methods further, and the paper urges the unlearning community to adopt editing baselines and treat refusal quality as an evaluation axis.

What carries the argument

The load-bearing mechanism is Equation 3, the editing-as-unlearning identity: replace each forget-query answer with a common refusal target $\emptyset$ and run an editing algorithm, so the model learns the mapping query $\to$ refusal while preserving other outputs. The empirical engine is the pair of lifelong editing methods WISE, which stores edits in a dynamic parametric side memory, and AlphaEdit, which constrains parameter updates to the null space to avoid side effects; they are supported by two adaptations: self-improvement, which uses the model's own in-context learning to craft tailored, entity-aware refusal answers, and query merging, which concatenates several forget queries into one edit with a single refusal target so that locate-and-edit methods like ROME and MEMIT can handle long unlearning sequences.

What would settle it

Run the paper's setting with WISE and AlphaEdit on a factual forget set, then probe the unlearned model's hidden activations on the original query (for example, train a linear probe on the model's own hidden states to decode the target answer, or measure the conditional probability of the correct answer when the refusal is bypassed by a paraphrased or 'continue' prompt). If the correct fact can still be decoded or elicited at high accuracy, the behavioral equivalence between refusal and erasure—and with it the claim that editing methods are unlearning methods—collapses.

Watch

Extended reading notes

Core claim

The paper's central claim is that knowledge editing and LLM unlearning are not separate tasks: unlearning is the special case of editing in which the target answer is replaced by a refusal, the 'empty set' $\emptyset$ of generative answers. Writing the editing objective as $f_{\Theta^*}(x)=\emptyset$ for forget queries and $f_{\Theta_0}(x)$ otherwise (Equation 3) turns any editing method into an unlearning method. On that basis the paper evaluates five editing methods (ROME, MEMIT, GRACE, WISE, AlphaEdit) against four unlearning baselines (GA, GD, KL, DPO) on TOFU's factual set (pretrained knowledge) and PISTOL (finetuned knowledge). It reports that WISE and AlphaEdit are the strongest editing methods and are competitive or superior to unlearning baselines, especially on pretrained knowledge, and that they produce refusal answers that human raters find more trustworthy and semantically entailed. The paper also claims that its self-improvement pipeline and query merging technique extend editing methods' reach, making even ROME and MEMIT perform well on long unlearning sequences.

Load-bearing premise

The assumption that carries the argument is that a model has unlearned when it refuses to answer a forget query, measured only on outputs; the paper never checks whether the underlying fact was actually removed from the model's parameters.

Editorial extensions

If this is right

  • Unlearning benchmarks should list WISE and AlphaEdit as standard baselines, since the paper shows they hold the forget/retain trade-off on pretrained factual knowledge.
  • Refusal quality should become a reported unlearning metric: pretrained-knowledge experiments in the paper show editing methods produce refusal answers rated higher on trustworthiness and semantic entailment than gradient-based unlearning.
  • The self-improvement pipeline implies that the editing target for unlearning can be synthesized by the model itself, improving rephrase-attack generalization at the cost of an alignment tax for WISE.
  • Query merging shows that ROME and MEMIT, previously ineffective under long sequences, can become competitive unlearning methods when several forget queries share one refusal target.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The authors' behavioral definition of forgetting—refusal at the output—does not establish that the fact is gone from the weights; a natural next test is probing hidden states of an 'unlearned' model to see whether the correct answer can still be decoded.
  • The refusal-as-$\emptyset$ equivalence suggests unlearning evaluation will eventually need to separate 'refusal to say' from 'genuine absence of knowledge,' because privacy and safety goals differ: a model that can be prompted or jailbroken back to the fact has not been unlearned in a legal sense.
  • Because many forget queries naturally share one refusal target, query merging could be generalized to group-level or topic-level unlearning, with the expected trade-off that longer merged contexts strain the underlying editing method.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 8 minor

Summary. This paper proposes to view LLM unlearning as a special case of knowledge editing in which the target answer is replaced by a refusal or 'empty set' response (Eq. 3). The authors evaluate five knowledge-editing methods (ROME, MEMIT, GRACE, WISE, AlphaEdit) against four unlearning methods (GA, GD, KL, DPO) on two datasets (TOFU factual/world knowledge for pretrained knowledge; PISTOL for finetuned knowledge) and three base models (Llama2-7B, Mistral-7B, Llama3.1-8B). They report that WISE and AlphaEdit are particularly strong unlearning baselines for pretrained knowledge, that editing methods produce more human-aligned refusal answers, and they propose two practical recipes: a self-improvement pipeline that uses the LLM itself to generate tailored refusal targets, and a query-merging technique that enables ROME and MEMIT to handle long forget sequences. The paper advocates that the unlearning community adopt state-of-the-art editing methods as baselines.

Significance. If the central claim is accepted, the paper offers a low-cost bridge between the knowledge-editing and unlearning communities: editing methods are mature and lightweight and could be plugged into unlearning benchmarks with minimal adaptation. The paper is a useful empirical contribution in that it compares a broader set of editing methods than previous work, evaluates on both pretrained and finetuned knowledge, and provides two concrete, reproducible recipes. The inclusion of human evaluation of refusal quality is a strength. However, the significance is limited by two factors: the evaluation measures behavioral refusal on the original query distribution rather than parametric erasure, and the paper's own rephrase-attack results show that the edited models still recall a substantial fraction of the target facts under rephrasing. As a result, the paper's headline recommendation that editing methods should be adopted as unlearning baselines is currently supported only under a narrow definition of unlearning; the significance would be substantially strengthened by evidence on knowledge retention or by an explicit reframing of the claim.

major comments (3)
  1. [3.1 and 4.1 (Eq. 3; Table 2)] The paper operationalizes unlearning purely behaviorally: Eq. (3) maps each forget query to a fixed refusal output, and the evaluation in Section 4.1 measures only output-surface metrics (Rouge1, Prob, MRR, Hit-Rate) on the original queries. No experiment checks whether the parametric knowledge is actually removed, despite the abstract's claim of "removing information from LLMs." The paper's own rephrase-attack results in Table 2 show residual recall: on the Llama2-7B factual dataset, WISE's forget-set Rouge1 increases from 0.28 on the original queries to 0.46 under GPT-4 rephrasings, and AlphaEdit's from 0.08 to 0.14. This indicates that the target facts remain retrievable when the query is rephrased and that the editing methods have learned to refuse on the training query distribution rather than to erase the underlying knowledge. This is structurally expected because WISE stores edits in side memory without modifying base parameters and AlphaEdit constrains updates to the null space to preserve existing knowledge. Since the central recommendation is that unlearning benchmarks should adopt editing methods as baselines, this recommendation currently rewards answer suppression on a narrow query distribution. Please either (i) explicitly reframe the claim as "strong baselines for producing refusal answers on the forget query distribution" and add a limitations paragraph discussing the distinction between behavioral refusal and parametric erasure, or (ii) add experiments that directly probe knowledge retention (e.g., diverse paraphrase attacks, internal representation probes, or counterfactual variants) and report those results.
  2. [Table 1; Appendix B.4] The comparative ranking that supports the central claim (e.g., Ob2 and Ob4) is reported without error bars, multiple seeds, or significance tests. Every cell in Table 1 and every point in Figures 3, 4, and 6 is a single run, so the reader cannot assess whether the differences between WISE/AlphaEdit and the unlearning baselines are meaningful. In addition, the hyperparameters for the editing methods are hand-tuned: Appendix B.4 states that mom2_n_samples for MEMIT and AlphaEdit is reduced to 1000 "considering computational resource constraints," while the unlearning baselines use learning rates taken from PISTOL. These choices are asymmetric and could affect the ranking. Please provide variance estimates (e.g., multiple seeds or random forget-set subsets) and justify the hyperparameter settings, or temper the comparative claims accordingly.
  3. [Table 3] The Llama3.1-8B results in Table 3 are mixed and are not discussed in the main text. On the factual dataset, WISE achieves the best forget Rouge1 (0.367) but the retain Rouge1 is only 0.592, well below GD (0.938) and AlphaEdit (0.847); AlphaEdit's forget Rouge1 is 0.517, which is worse than WISE's but still better than DPO's 0.45. The text's claim that "WISE and AlphaEdit consistently perform well across all settings" (Ob1) and that editing methods are strong for pretrained knowledge does not hold cleanly on this third model. Since the paper argues for a general recommendation to the unlearning community, please analyze these results explicitly and state the scope over which the claim is intended to apply.
minor comments (8)
  1. [3.1] The notation ∅ ("empty set") is used to denote a refusal sentence, which is mathematically a set rather than a string; please define a different symbol (e.g., r(x)) or clarify the abuse of notation.
  2. [Figure 1] In the provided text, the figure contains a garbled fragment ("archivihtml") and the text in panels is very small; please ensure the final figure is rendered legibly.
  3. [Appendix Table 4] In the WISE row, the value "0087" appears to be a typo for "0.087"; please fix.
  4. [4.4 / Figure 6 (left)] The human evaluation relies on only 20 participants and reports no inter-annotator agreement or significance tests; given the small differences (e.g., WISE trustworthiness 2.56 to 2.38, AlphaEdit entailment 1.55 to 1.73), please present these results as qualitative or add statistical support.
  5. [3.2 / Figure 6 (right)] Please specify how the merged queries are concatenated (e.g., separator tokens, instruction template) and how the single refusal target is generated, for reproducibility; the paper currently states only that several queries are concatenated into one.
  6. [4.2] The comparison is not fully controlled because GD and KL use the retain set while editing methods do not; the paper notes this at the end of Section 4.1, but the potential advantage to editing methods is not discussed. A brief discussion would help readers interpret the comparative rankings.
  7. [4.4 (Obs6)] The parenthetical "merging 5 queries into 1 (16 samples after merging)" is confusing; it would be clearer to write "80 forget samples, 80/5 = 16 editing samples."
  8. [Limitations] The Limitations section is very brief and does not mention the behavioral-versus-parametric distinction or the mixed Llama3.1 results; please expand it accordingly.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the evaluation is empirical and self-contained; self-citations to WISE and PISTOL are not load-bearing, and the unlearning-as-refusal framing is an operationalization rather than a result derived from itself.

full rationale

The paper's central claim is an empirical benchmark comparison, not a derivation. Equation 3 operationalizes unlearning as editing to a refusal target, but this is a stated modeling choice, and the conclusion that WISE and AlphaEdit are strong baselines rests on measurements in Tables 1-3, not on Eq. 3 alone. No parameter is fitted and then renamed a prediction; the self-improvement pipeline is evaluated by independent human raters, and query merging is an input transformation whose effects are measured. Self-citations to WISE and PISTOL are present, but the methods are executed and benchmarked rather than cited as authority for the conclusion, so they are not load-bearing. The definitional choice to treat refusal as empty-set knowledge is contestable, and the paper does not demonstrate parametric erasure; however, that is a correctness and scope concern, not a circular reduction. The paper therefore has no circular step that reduces its result to its inputs.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper's central claim does not rest on a mathematical derivation; it rests on empirical comparisons. The main hidden costs are hand-chosen hyperparameters (Appendix B.4), selection of the merged-query count on the evaluation set (Figure 6), and benchmark assumptions that TOFU factual QA pairs represent pretrained knowledge and that refusal behavior is equivalent to unlearning. No new physical or architectural entities are introduced.

free parameters (3)
  • learning rates for unlearning baselines = 2e-5 for GA/GD/KL, 1.5e-5 for DPO
    Taken from PISTOL and set by hand; the comparison between editing and unlearning methods depends on these hyperparameters (Appendix B.4).
  • mom2_n_samples for MEMIT and AlphaEdit = 1000 (reduced from default; default for ROME/GRACE/WISE)
    The covariance sample count is reduced for computational reasons, which can affect the stability and performance of MEMIT and AlphaEdit weight updates (Appendix B.4).
  • merged query count in query merging = 5 (swept 2-10; best on the same forget set)
    The number of merged queries is swept and the peak (16 merged samples for 80 forget samples) is highlighted, which is selection on the evaluation set (Section 4.4, Figure 6).
assumptions (3)
  • domain assumption The TOFU factual QA pairs are treated as knowledge the model acquired during pretraining.
    The paper calls this dataset 'pretrained knowledge' based on high pre-unlearning Rouge1 scores (0.82 for Llama2-7B, 0.86 for Mistral-7B), but actual membership in the pretraining corpus is not verified (Section 4.1, 4.2).
  • domain assumption Behavioral refusal to answer is an acceptable operationalization of unlearning.
    Equation (3) replaces the forgotten answer with an empty-set refusal and evaluates only model outputs; the paper does not test whether the parametric knowledge itself was erased, and related work cited in the paper notes that unlearning often obscures rather than removes data.
  • ad hoc to paper A single refusal answer is a valid editing target for multiple merged queries.
    Query merging concatenates several forget questions into one editing sample with one shared refusal target, assuming the same refusal is appropriate for all of them and that the editing method preserves reliability (Section 3.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Editing as Unlearning: Are Knowledge Editing Methods Strong Baselines for Large Language Model Unlearning?." pith.science (2026). https://pith.science/paper/LF6ZWSGN

@misc{pith2026250519855,
  author       = {Pith},
  title        = {Pith review of: Editing as Unlearning: Are Knowledge Editing Methods Strong Baselines for Large Language Model Unlearning?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LF6ZWSGN}},
  note         = {Machine review of arXiv:2505.19855}
}
abstract

Large language Model (LLM) unlearning, i.e., selectively removing information from LLMs, is vital for responsible model deployment. Differently, LLM knowledge editing aims to modify LLM knowledge instead of removing it. Though editing and unlearning seem to be two distinct tasks, we find there is a tight connection between them. In this paper, we conceptualize unlearning as a special case of editing where information is modified to a refusal or "empty set" $\emptyset$ response, signifying its removal. This paper thus investigates if knowledge editing techniques are strong baselines for LLM unlearning. We evaluate state-of-the-art (SOTA) editing methods (e.g., ROME, MEMIT, GRACE, WISE, and AlphaEdit) against existing unlearning approaches on pretrained and finetuned knowledge. Results show certain editing methods, notably WISE and AlphaEdit, are effective unlearning baselines, especially for pretrained knowledge, and excel in generating human-aligned refusal answers. To better adapt editing methods for unlearning applications, we propose practical recipes including self-improvement and query merging. The former leverages the LLM's own in-context learning ability to craft a more human-aligned unlearning target, and the latter enables ROME and MEMIT to perform well in unlearning longer sample sequences. We advocate for the unlearning community to adopt SOTA editing methods as baselines and explore unlearning from an editing perspective for more holistic LLM memory control.

Figures

Figures reproduced from arXiv: 2505.19855 by the authors.

Figure 1
Figure 1. Illustrations of the connection between editing and unlearning for LLMs. A: Editing aims to alter the knowledge to a target. B: Unlearning tries to remove the knowledge and generate an "empty" (without information) answer. C: Editing as unlearning, can be done by editing that alters the knowledge into a target refusal answer. unlearning focuses on removing the knowledge, knowledge editing aims to alter the knowledge… view at source ↗
Figure 2
Figure 2. Methods of improving editing algorithms in unlearning settings. A: Self-improvement pipeline improves generalization and human value alignment for AlphaEdit and WISE. B: Query merging technique enables ROME and MEMIT to perform well under long unlearning sequences. Equation 3 bridges from editing to unlearning, making it applicable to verify whether editing meth￾ods are strong baselines for unlearning. 3.2 Improving… view at source ↗
Figure 3
Figure 3. Results of different numbers of forget sam￾ples. Factual dataset, Llama2-7B. 4.3 Comprehensive Analysis (RQ1 & RQ2) We study the capabilities of editing methods under rephrase attack and different numbers of forget samples. We note that the rephrase attack is noted as the generalization metric in knowledge editing (Wang et al., 2024a), and we use GPT-4 to synthesize the rephrased queries. For the figures, to get a m… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comprehensive analysis of unlearning performances. The same setting as [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Case study of LLMs’ answers after unlearn￾ing. Factual dataset, Llama2-7B. 4.4 Improving Editing Methods in Unlearning Settings (RQ3) LLM outputs should align with human val￾ues (Wang et al., 2023). However, we observe that some unlearning methods cause models to gener…
Figure 6
Figure 6. Figure 6: Results of improving editing in unlearning. Factual dataset, Llama2-7B. Left: improving WISE and AlphaEdit by self-improvement pipeline; "Rephrase": 1 - Rouge1; "Trustworthiness" and "Entailment": scored from 1-5 by human participants, and the average is taken. Right: …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. BalDRO: A Distributionally Robust Optimization based Framework for Large Language Model Unlearning

    cs.LG 2026-01 conditional novelty 5.0 of 10

    BalDRO makes LLM unlearning more balanced by updating against a worst-case-weighted forget distribution, improving forget quality on TOFU/MUSE at stable utility.

Reference graph

Works this paper leans on

56 extracted references · 29 canonical work pages · cited by 1 Pith paper

  1. [1]

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

  2. [2]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877--1901

  3. [3]

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

  4. [4]

    Hoyeon Chang, Jinho Park, Seonghyeon Ye, Sohee Yang, Youngkyung Seo, Du-Seong Chang, and Minjoon Seo. 2024. How do large language models acquire factual knowledge during pretraining? In The Thirty-eighth Annual Conference on Neural Information Processing Systems

  5. [5]

    Badhan Chandra Das, M Hadi Amini, and Yanzhao Wu. 2025. Security and privacy challenges of large language models: A survey. ACM Computing Surveys, 57(6):1--39

  6. [6]

    Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Shi Jie, Xiang Wang, Xiangnan He, and Tat-Seng Chua. 2025. Alphaedit: Null-space constrained knowledge editing for language models. In The Thirteenth International Conference on Learning Representations

  7. [7]

    Aditya Golatkar, Alessandro Achille, and Stefano Soatto. 2020. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9304--9312

  8. [8]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

Show all 56 references
  1. [9]

    Phillip Guo, Aaquib Syed, Abhay Sheshadri, Aidan Ewart, and Gintare Karolina Dziugaite. 2024. Mechanistic unlearning: Robust knowledge unlearning and editing via mechanistic localization. arXiv preprint arXiv:2410.12949

  2. [10]

    Tom Hartvigsen, Swami Sankaranarayanan, Hamid Palangi, Yoon Kim, and Marzyeh Ghassemi. 2023. Aging with grace: Lifelong model editing with discrete key-value adaptors. Advances in Neural Information Processing Systems, 36:47934--47959

  3. [11]

    Shariqah Hossain. 2025. Investigating Model Editing for Unlearning in Large Language Models. Ph.D. thesis, Massachusetts Institute of Technology

  4. [12]

    James Y Huang, Wenxuan Zhou, Fei Wang, Fred Morstatter, Sheng Zhang, Hoifung Poon, and Muhao Chen. 2024. Offset unlearning for large language models. arXiv preprint arXiv:2404.11045

  5. [13]

    Jiabao Ji, Yujian Liu, Yang Zhang, Gaowen Liu, Ramana Kompella, Sijia Liu, and Shiyu Chang. 2024. Reversing the forget-retain objectives: An efficient llm unlearning framework from logit difference. Advances in Neural Information Processing Systems, 37:12581--12611

  6. [14]

    Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, and 1 others. 2024. Mixtral of experts. arXiv preprint arXiv:2401.04088

  7. [15]

    Houcheng Jiang, Junfeng Fang, Ningyu Zhang, Guojun Ma, Mingyang Wan, Xiang Wang, Xiangnan He, and Tat-seng Chua. 2025. Anyedit: Edit any knowledge encoded in language models. arXiv preprint arXiv:2502.05628

  8. [16]

    Kevin Kuo, Amrith Setlur, Kartik Srinivas, Aditi Raghunathan, and Virginia Smith. 2025. Exact unlearning of finetuning data via model merging at scale. arXiv preprint arXiv:2504.04626

  9. [17]

    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, Gabriel Mukobi, and 1 others. The wmdp benchmark: Measuring and reducing malicious use with unlearning. In Forty-first International Conferen...

  10. [18]

    Zhoubo Li, Ningyu Zhang, Yunzhi Yao, Mengru Wang, Xi Chen, and Huajun Chen. 2024. Unveiling the pitfalls of knowledge editing for large language models. In The Twelfth International Conference on Learning Representations

  11. [19]

    Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai, Daya Guo, and 1 others. 2024 a . Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434

  12. [20]

    Bo Liu, Qiang Liu, and Peter Stone. 2022. Continual learning and private unlearning. In Conference on Lifelong Learning Agents, pages 243--254. PMLR

  13. [21]

    Chris Liu, Yaxuan Wang, Jeffrey Flanigan, and Yang Liu. 2024 b . Large language model unlearning via embedding-corrupted prompts. Advances in Neural Information Processing Systems, 37:118198--118266

  14. [22]

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

  15. [23]

    Xinbei Ma, Tianjie Ju, Jiyang Qiu, Zhuosheng Zhang, Yulong Wang, and 1 others. 2024. Is it possible to edit large language models robustly? In ICLR 2024 Workshop on Large Language Model (LLM) Agents

  16. [24]

    Tofu: A task of fictitious unlearning for llms

    Pratyush Maini, Zhili Feng, Avi Schwarzschild, Zachary Chase Lipton, and J Zico Kolter. Tofu: A task of fictitious unlearning for llms. In First Conference on Language Modeling

  17. [25]

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in neural information processing systems, 35:17359--17372

  18. [26]

    Kevin Meng, Arnab Sen Sharma, Alex J Andonian, Yonatan Belinkov, and David Bau. 2023. Mass-editing memory in a transformer. In The Eleventh International Conference on Learning Representations

  19. [27]

    Fast model editing at scale

    Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. Fast model editing at scale. In International Conference on Learning Representations

  20. [28]

    Quoc Phong Nguyen, Bryan Kian Hsiang Low, and Patrick Jaillet. 2020. Variational bayesian unlearning. Advances in Neural Information Processing Systems, 33:16025--16036

  21. [29]

    Jasmine Chiat Ling Ong, Shelley Yin-Hsi Chang, Wasswa William, Atul J Butte, Nigam H Shah, Lita Sui Tjien Chew, Nan Liu, Finale Doshi-Velez, Wei Lu, Julian Savulescu, and 1 others. 2024. Ethical and regulatory challenges of large language models in medicine. The Lancet Digital...

  22. [30]

    Stuart L Pardau. 2018. The california consumer privacy act: Towards a european-style privacy regime in the united states. J. Tech. L. & Pol'y, 23:68

  23. [31]

    Xinchi Qiu, William F Shen, Yihong Chen, Nicola Cancedda, Pontus Stenetorp, and Nicholas D Lane. 2024. Pistol: Dataset compilation pipeline for structural unlearning of llms. arXiv preprint arXiv:2406.16810

  24. [32]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728--53741

  25. [33]

    William F Shen, Xinchi Qiu, Meghdad Kurmanji, Alex Iacob, Lorenzo Sani, Yihong Chen, Nicola Cancedda, and Nicholas D Lane. 2025. Lunar: Llm unlearning via neural activation redirection. arXiv preprint arXiv:2502.07218

  26. [34]

    Chenmien Tan, Ge Zhang, and Jie Fu. 2024. Massive editing for large language models via meta learning. In The Twelfth International Conference on Learning Representations

  27. [35]

    Arun James Thirunavukarasu, Darren Shu Jeng Ting, Kabilan Elangovan, Laura Gutierrez, Ting Fang Tan, and Daniel Shu Wei Ting. 2023. Large language models in medicine. Nature medicine, 29(8):1930--1940

  28. [36]

    Bozhong Tian, Xiaozhuan Liang, Siyuan Cheng, Qingbin Liu, Mengru Wang, Dianbo Sui, Xi Chen, Huajun Chen, and Ningyu Zhang. 2024. To forget or not? towards practical knowledge unlearning for large language models. In Findings of the Association for Computational Linguistics: EM...

  29. [37]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, and 1 others. 2023 a . Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

  30. [38]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, and 1 others. 2023 b . Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

  31. [39]

    Akshaj Kumar Veldanda, Shi-Xiong Zhang, Anirban Das, Supriyo Chakraborty, Stephen Rawls, Sambit Sahu, and Milind Naphade. 2024. Llm surgery: Efficient knowledge unlearning and editing in large language models. arXiv preprint arXiv:2409.13054

  32. [40]

    Paul Voigt and Axel Von dem Bussche. 2017. The eu general data protection regulation (gdpr). A practical guide, 1st ed., Cham: Springer International Publishing, 10(3152676):10--5555

  33. [41]

    Boxin Wang, Weixin Chen, Hengzhi Pei, Chulin Xie, Mintong Kang, Chenhui Zhang, Chejian Xu, Zidi Xiong, Ritik Dutta, Rylan Schaeffer, and 1 others. 2023. Decodingtrust: A comprehensive assessment of trustworthiness in gpt models. In NeurIPS

  34. [42]

    Junyang Wang, Haiyang Xu, Haitao Jia, Xi Zhang, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang. 2025 a . Mobile-agent-v2: Mobile device operation assistant with effective navigation via multi-agent collaboration. Advances in Neural Information Processing Systems, ...

  35. [43]

    Lingzhi Wang, Xingshan Zeng, Jinsong Guo, Kam-Fai Wong, and Georg Gottlob. 2025 b . Selective forgetting: Advancing machine unlearning techniques and evaluation in language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 843--851

  36. [44]

    Peng Wang, Zexi Li, Ningyu Zhang, Ziwen Xu, Yunzhi Yao, Yong Jiang, Pengjun Xie, Fei Huang, and Huajun Chen. 2024 a . Wise: Rethinking the knowledge memory for lifelong model editing of large language models. Advances in Neural Information Processing Systems, 37:53764--53797

  37. [45]

    Peng Wang, Ningyu Zhang, Bozhong Tian, Zekun Xi, Yunzhi Yao, Ziwen Xu, Mengru Wang, Shengyu Mao, Xiaohan Wang, Siyuan Cheng, and 1 others. 2024 b . Easyedit: An easy-to-use knowledge editing framework for large language models. In Proceedings of the 62nd Annual Meeting of the ...

  38. [46]

    Yaxuan Wang, Jiaheng Wei, Chris Yuhao Liu, Jinlong Pang, Quan Liu, Ankit Parag Shah, Yujia Bao, Yang Liu, and Wei Wei. 2025 c . Llm unlearning via loss adjustment with only forget data. ICLR

  39. [47]

    Yuxiang Wei, Chunqiu Steven Xia, and Lingming Zhang. 2023. Copiloting the copilots: Fusing large language models with completion engines for automated program repair. In Proceedings of the 31st ACM Joint European Software Engineering Conference and Symposium on the Foundations...

  40. [48]

    Haoming Xu, Ningyuan Zhao, Liming Yang, Sendong Zhao, Shumin Deng, Mengru Wang, Bryan Hooi, Nay Oo, Huajun Chen, and Ningyu Zhang. 2025. Relearn: Unlearning via learning for large language models. arXiv preprint arXiv:2502.11190

  41. [49]

    Jin Yao, Eli Chien, Minxin Du, Xinyao Niu, Tianhao Wang, Zezhou Cheng, and Xiang Yue. 2024 a . Machine unlearning of pre-trained large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8403--8419

  42. [50]

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

  43. [51]

    Yunzhi Yao, Peng Wang, Bozhong Tian, Siyuan Cheng, Zhoubo Li, Shumin Deng, Huajun Chen, and Ningyu Zhang. 2023. Editing large language models: Problems, methods, and opportunities. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages...

  44. [52]

    Lang Yu, Qin Chen, Jie Zhou, and Liang He. 2024. Melo: Enhancing model editing with neuron-indexed dynamic lora. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19449--19457

  45. [53]

    Binchi Zhang, Zhengzhang Chen, Zaiyi Zheng, Jundong Li, and Haifeng Chen. 2025. Resolving editing-unlearning conflicts: A knowledge codebook framework for large language model updating. arXiv preprint arXiv:2502.00158

  46. [54]

    Jiamu Zheng, Jinghuai Zhang, Tianyu Du, Xuhong Zhang, Jianwei Yin, and Tao Lin. 2025. Collabedit: Towards non-destructive collaborative knowledge editing. In The Thirteenth International Conference on Learning Representations

  47. [55]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  48. [56]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.