Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

Not Every Token Needs Forgetting: Selective Unlearning to Limit Change in Utility in Large Language Model Unlearning

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Selective Unlearning forgets only the tokens whose prediction scores diverge between two assistant models, matching full-token unlearning on forget quality while preserving retained knowledge on TOFU and MUSE-News.

desk verdict The core idea is a real but modest extension of token-level selection, yet the reported utility gains are invalidated by threshold tuning on the evaluation retain metrics and contradicted by the paper's own Table 2. read the letter →

arxiv 2506.00876 v1 pith:2AKZAKPC submitted 2025-06-01 cs.CL

Yixin Wan , Anil Ramakrishna , Kai-Wei Chang , Volkan Cevher , Rahul Gupta This is my paper
classification cs.CL
keywords LLMunlearningselectivetoken-levelselectionutilitypreservationgradientascentn-gramlanguagemodelsTOFUbenchmarkMUSE
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

Selective Unlearning (SU) claims that an LLM forgets targeted data more safely when the unlearning loss is applied only to tokens carrying forget-specific information, rather than to every token in the target document. The paper identifies those tokens by training two small assistant models on different data splits and thresholding the difference in their next-token prediction scores. On the TOFU and MUSE-News benchmarks, SU matches the forget quality of conventional gradient-ascent unlearning while retaining markedly more knowledge from the retain set, real-world facts, and real-author facts. If correct, this decouples forget effectiveness from utility loss, making unlearning practical for deployment where a model must remain broadly useful after removing sensitive content.

What carries the argument

The central object is the SU selection criterion $S(t_i) = 1$ iff $|p^1_\theta(t_i|t_{<i}) - p^2_\theta(t_i|t_{<i})| > \gamma$, where the two assistant models are trained on different data splits so that their disagreement marks tokens whose predictability depends on forget-set content. This selection function is the mechanism that decides which tokens receive the unlearning loss; the actual forgetting is done by gradient ascent on a five-token window around each selected token, paired with gradient descent on the retain set. The paper implements the assistants either as small n-gram models (about 20 MB) or as fine-tuned Mistral-7B models.

What would settle it

Take a synthetic forget document in which a rare named entity is the only piece of targeted knowledge while surrounding words are generic, train the two assistant models on full versus retain data, and check whether SU's threshold selects the entity token and not function words; if the selected tokens are not the ones whose removal changes answers to forget probes, the mechanism fails. More directly, compare SU against random token selection of the same size: if random subsets match SU's retain utility, the utility gain comes from doing less unlearning, not from the selection criterion.

Watch

Extended reading notes

Core claim

SU's central claim is that "not every token needs forgetting": the model parameters should be updated to unlearn only a critical subset of tokens within the forget set, while common tokens such as pronouns and prepositions keep their learned representations. To find that subset, SU trains two assistant models—one on full data and one on retain data (or other split combinations)—and selects a token $t_i$ when $|p^1_\theta(t_i|t_{<i}) - p^2_\theta(t_i|t_{<i})| > \gamma$. Unlearning then maximizes the language-model loss on a five-token window around each selected token, and minimizes it on the retain set. The paper reports that on TOFU and MUSE-News, SU with either n-gram or LLM assistants achieves forget scores comparable to full-token unlearning baselines while reaching the highest retain utility across all compared methods.

Load-bearing premise

The load-bearing premise is that the size of the disagreement between two assistant models' token predictions points exactly at tokens carrying forget-set-specific information—and if that disagreement mostly reflects ordinary word-frequency differences, the selection will miss the target knowledge or erase generic tokens.

Editorial extensions

If this is right

  • SU reaches forget scores comparable to full-token unlearning on both benchmarks, so a deployment can still expect targeted content to be removed.
  • SU keeps retain-set knowledge, real-world knowledge, and real-author knowledge measurably higher than the six compared baselines, so an unlearned model remains useful beyond the narrow forget set.
  • N-gram-based SU achieves the best overall utility and uses only about 20 MB of assistant-model memory, so token-level selection is practical for moderate-scale unlearning runs.
  • The extra cost of SU is confined to training the two assistant models; after unlearning, selection assistants are no longer needed, so inference latency and memory are unaffected.

Reading between the lines

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

  • The paper never inspects which tokens are selected; an independent analysis could test whether selection really concentrates on names and rare content words, or whether it often picks functional tokens that merely differ in frequency between splits.
  • Because the selection threshold is tuned on evaluation data, an open question is how the method would behave with a fixed, unsupervised threshold or a per-document adaptive one; if the threshold is fragile, utility gains may shrink outside benchmark settings.
  • A random-subset control of the same token count would isolate whether SU's utility benefit comes from the selection criterion or simply from unlearning fewer tokens; the current experiments do not include this control.
  • In real deployments the original training data may not be available to train the "full data" assistant; a practical variant might use the model itself or a proxy corpus, but the paper does not test that setting.
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

4 major / 4 minor

Summary. The paper proposes Selective Unlearning (SU), a token-level unlearning method for LLMs that computes a score difference between two assistant models trained on different data splits and applies the unlearning loss only to tokens whose score difference exceeds a threshold γ. The authors claim that SU achieves forget quality comparable to full-token unlearning methods while substantially preserving retain-set utility. Experiments are conducted on TOFU and MUSE-News against six baseline unlearning algorithms, with two assistant model types (n-gram and LLM-based). The central claim is that token-selective unlearning can decouple forget quality from retain utility loss.

Significance. If the central claim were valid, SU would address a genuine problem in LLM unlearning: indiscriminate gradient ascent on all tokens in a forget document degrades general knowledge encoded in common tokens. The idea of using a lightweight n-gram assistant model for token selection is interesting and cost-effective, and the paper includes a reasonably broad set of baselines and two standard benchmarks. However, the empirical evidence as presented does not support the utility-preservation claim: the results are internally contradicted by the paper's own Table 2, and the selection threshold is tuned on the evaluation metrics themselves. The paper does provide memory-cost measurements for the n-gram assistants and some qualitative examples, but these do not outweigh the evaluation-protocol and internal-consistency issues.

major comments (4)
  1. [Section 4.4 and Table 2] The claim that 'both SU methods achieve better knowledge memorization on MUSE-News' retain set' is directly contradicted by Table 2. On KnowMem(Retain), SU(N-Gram)=0.20 and SU(LLM)=0.19 are lower than NPO+GD=0.30, GA+KL=0.26, and NPO+KL=0.25. On TOFU Truth(Retain), SU(LLM)=0.48 only ties NPO+GD=0.48 and SU(N-Gram)=0.44 trails both NPO+GD and NPO+KL=0.45. GA+GD also exceeds SU on Truth(Real-World) (0.73 vs 0.62 for SU(N-Gram)) and Truth(Real-Author) (0.75 vs 0.72). The abstract's claim that SU 'significantly preserves the model's utility in the retaining set' is therefore not supported by the paper's own numbers.
  2. [Section 4.3 and Appendix B.2] The selection threshold γ is chosen by searching over {0.2, 0.5, 0.8, 0.9} and selecting the value that maximizes performance on the retain set, exactly the same retain-set metrics (KnowMem(Retain), Truth(Retain)) that Table 2 reports as SU's utility advantage. No equivalent threshold tuning is described for any of the six baselines. This is selection on the evaluation data: the reported SU utility numbers are post-hoc optimized values, not model predictions, so the comparison against untuned baselines is unfair. The central claim that SU preserves utility is thus not established by these experiments.
  3. [Section 3.1, Eq. (1), and Appendix B.2] The load-bearing premise that |p1(t_i|t_<i) − p2(t_i|t_<i)| > γ identifies tokens carrying forget-specific information is not validated anywhere in the paper. No analysis is given of which tokens are selected (e.g., part-of-speech distribution, overlap with named entities, content-word versus stopword fractions), so the reader cannot tell whether the score divergence reflects forget-specific knowledge or ordinary distributional differences between the two assistant training corpora. Because γ itself is tuned on retain performance, the selection mechanism's link to forget-specificity is circular: the threshold is chosen to maximize retain utility, and then retain utility is reported as the method's benefit. A concrete remedy would be to fix γ a priori or tune it on a held-out validation split, and to report token-selection statistics that demonstrate the selected tokens are indeed forget-related.
  4. [Algorithm 1] The bookkeeping of the unlearning loss token list `l` is under-specified and inconsistent with the description that 'we calculate the unlearning loss for 5-grams surrounding each selected token.' When two selected tokens are within four positions of each other, the same index can be added to `l` multiple times, and the `else if i in l` branch removes only one occurrence, so overlapping windows cause some tokens to receive disproportionately large weight in the loss. Moreover, a non-selected token to the right of a selected token is removed when its own index arrives, so the effective window is asymmetric (left context survives, right context is pruned) unless re-added by a later selected token. As written, the algorithm does not guarantee a 5-gram window and is not sufficient for reproducible implementation.
minor comments (4)
  1. [Section 3.2] The word 'interoperability' in 'due to their efficiency and interoperability' appears to be a typo for 'interpretability' or 'interpretability' in the sense of transparency; please clarify.
  2. [Section 4.4] The statement 'SU with N-Gram-based selection mechanism achieves the overall best result' is not supported by Table 2: SU(N-Gram) has lower Truth(Retain) than SU(LLM) (0.44 vs 0.48), and on MUSE KnowMem(Retain) the difference (0.20 vs 0.19) is within the noise one might expect without reported variance.
  3. [Table 2] The arrows in the column headers (e.g., 'VerbMem (↓0)') are ambiguous because the text labels in the table do not always align with the metric descriptions in Appendix A; please clarify in the table caption which direction indicates better performance for each column.
  4. [Appendix D, Table 5] The SU(N-Gram) generation '7 said said said said noreferrer noreferrer the said...' suggests that the model may have lost fluency or coherence; since utility preservation should include general language quality, the paper should discuss whether such outputs are acceptable and whether the model retains basic generation ability.

Circularity Check

1 steps flagged · score 6.0 of 10

Utility-preservation claim is fitted by gamma search on retain-set metrics; token-selection mechanism itself is not definitionally circular.

  1. fitted input called prediction [Section 4.3 (Token Selection) and Appendix B.2 (Hyper-Parameter Searching)]
    "To search for the best hyper-parameter for the SU method, we first experimented with three thresholds ... 0.2, 0.5, and 0.8. Figure 4 visualizes the result of ablation experiments... using one model trained on full data and one on retain data with a selection threshold of 0.8 achieves the best result. Based on the trend that we observe in experiments, we continued the search to experiment with an additional threshold of 0.9, which we eventually select for reporting experiment results."

    The SU-specific free parameter γ (Eq. 1) determines which tokens are unlearned, and Appendix B.2 reports that the final thresholds were selected by searching {0.2, 0.5, 0.8, 0.9} against model performance on the retain set—the same KnowMem(Retain) and Truth(Retain) metrics later reported in Table 2 as SU's utility advantage. Thus the headline utility-preservation result is a post-hoc choice of γ on the evaluation metric, not an independent prediction. No equivalent validation split or threshold search is described for the six baselines, so the comparison is not at equal selection effort; the 'significant' utility gain is partly forced by construction of the evaluation protocol.

full rationale

The core token-selection derivation is not definitionally circular: SU uses the score divergence between two assistant models to choose which tokens to unlearn, and forget-set quality is evaluated on forget metrics that are not used to fit γ. The circularity is concentrated in the utility-preservation claim. Appendix B.2 shows that the only SU-specific hyperparameter, γ, was chosen by searching over thresholds and picking the value that performed best on the retain-set metrics—exactly the metrics in Table 2 used to claim that SU 'significantly preserves' retain utility. This is selection on the evaluation data, so the reported utility advantage is partly an artifact of the tuning protocol rather than an independent result. Separately, Table 2 does not support the Section 4.4 statement that 'both SU methods achieve better knowledge memorization on MUSE-News’ retain set': NPO+GD scores 0.30 vs SU(N-Gram) 0.20 and SU(LLM) 0.19, and on TOFU Truth(Retain) SU(LLM) at 0.48 only ties NPO+GD. That is an internal-inconsistency/correctness concern, not a circularity. No load-bearing self-citation or imported uniqueness theorem is present; the assistant-model idea is independently implemented. Overall, the method has independent content, but the central utility claim is weakened by fitting γ on the reported evaluation metric, giving a partial-circularity score of 6.

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

The method relies on three domain assumptions: the two-assistant score difference isolates forget-specific tokens, the selected subset is sufficient for unlearning, and the benchmarks are valid proxies. The free parameters are the threshold and split choices, both tuned on the evaluation data, which undermines the independence of the reported utility results.

free parameters (3)
  • selection threshold gamma = 0.9 for TOFU (N-Gram and LLM), 0.8 for MUSE (both)
    Chosen by hyperparameter search over 0.2, 0.5, 0.8, then extended to 0.9 based on retain performance (Appendix B.2). Tuned on the evaluation data.
  • assistant split combination = TOFU (N-Gram): Full+Retain; TOFU (LLM): Full+Forget; MUSE (both): Full+Retain
    Selected after experiments based on which split/threshold yields best retain utility; a model selection choice that affects results.
  • n-gram order = 3-gram for TOFU, 5-gram for MUSE
    Chosen by authors; no ablation reported for this choice.
assumptions (3)
  • domain assumption The two assistant models trained on full vs retain data differ in their predictions mainly on tokens that encode forget-set-specific information.
    Section 3.1: SU places a threshold on score difference between the two assistants to identify forget-specific tokens; if the divergence is driven by general frequency shifts, selection is noisy.
  • domain assumption Unlearning only the selected tokens and their surrounding 5-grams removes the target knowledge while preserving retain utility.
    Section 3.1 and Algorithm 1: the method assumes the selected subset is sufficient for forgetting.
  • domain assumption The TOFU and MUSE benchmarks' forget/retain splits and metrics faithfully measure unlearning quality and utility.
    Section 4.1: evaluation follows Bu et al. (2024); no argument that these benchmarks capture real-world unlearning requirements.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Not Every Token Needs Forgetting: Selective Unlearning to Limit Change in Utility in Large Language Model Unlearning." pith.science (2026). https://pith.science/paper/2AKZAKPC

@misc{pith2026250600876,
  author       = {Pith},
  title        = {Pith review of: Not Every Token Needs Forgetting: Selective Unlearning to Limit Change in Utility in Large Language Model Unlearning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2AKZAKPC}},
  note         = {Machine review of arXiv:2506.00876}
}
read the original abstract

Large Language Model (LLM) unlearning has recently gained significant attention, driven by the need to remove unwanted information, such as private, sensitive, or copyrighted content, from LLMs. However, conventional unlearning approaches indiscriminately update model parameters to forget all tokens in a target document, including common tokens (e.g., pronouns, prepositions, general nouns) that carry general knowledge. In this paper, we highlight that not every token needs forgetting. We propose Selective Unlearning (SU), which identifies a critical subset of tokens within the forgetting set that is relevant to the unwanted information, and unlearns only those tokens. Experiments on two benchmarks and six baseline unlearning algorithms demonstrate that SU not only achieves effective unlearning on the targeted forget data, but also significantly preserves the model's utility in the retaining set.

Figures

Figures reproduced from arXiv: 2506.00876 by the authors.

Figure 1
Figure 1. Example of how tokens are selected for unlearning. Red blocks indicate unlearned tokens, on which the forget￾ting loss is calculated. SU avoids the forgetting of general information like “that”, therefore preserving model utility. troduce Selective Unlearning (SU), a novel frame￾work that utilizes two assistant models with differ￾ent scopes of knowledge to identify and unlearn only a subset of tokens that carry forg… view at source ↗
Figure 2
Figure 2. The proposed SU framework. We use 2 assis￾tant models, trained on different data splits, to facilitate the token selection process. Based on the difference between their prediction scores, we can choose to only unlearn tokens that contain information unique to the forget dataset. In the field of language model pre-training, Lin et al. (2024)’s work showed that not all tokens are needed for training a model. Specific… view at source ↗
Figure 3
Figure 3. Qualitative example of how SU excels at preserving utility on retain knowledge. 5 Conclusion In this paper, we introduce Selective Unlearning (SU), a novel framework that selectively erases es￾sential tokens with forget set-specific information, while keeping model knowledge on more common and universal tokens. Comprehensive experiments across two benchmarks and six baseline unlearning approaches demonstrated that S… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The influence of different selection thresholds [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. A Mechanistic Perspective and Circuit-Guided Difficulty Metric for Unlearning

    cs.LG 2026-01 conditional novelty 6.0 of 10

    A circuit-similarity score predicts which samples an LLM unlearning method will fail to erase, with hard samples relying on deeper, output-facing pathways.

  2. Maximizing Local Entropy Where It Matters: Prefix-Aware Localized LLM Unlearning

    cs.CL 2026-01 conditional novelty 6.0 of 10

    PALU shows that unlearning only needs local intervention—the first few tokens of the sensitive span and the top-k logits—not full-sequence, full-vocabulary suppression.

Reference graph

Works this paper leans on

25 extracted references · 7 canonical work pages · cited by 2 Pith papers

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  4. [4]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  5. [5]

    Brown, Vincent J

    Peter F. Brown, Vincent J. Della Pietra, Peter V. deSouza, Jenifer C. Lai, and Robert L. Mercer. 1992. https://aclanthology.org/J92-4003 Class-based n-gram models of natural language . Computational Linguistics, 18(4):467--480

    work page 1992

  6. [6]

    Zhiqi Bu, Xiaomeng Jin, Bhanukiran Vinzamuri, Anil Ramakrishna, Kai-Wei Chang, Volkan Cevher, and Mingyi Hong. 2024. https://arxiv.org/abs/2410.22086 Unlearning as multi-task optimization: A normalized gradient difference approach with an adaptive learning rate . Preprint, arXiv:2410.22086

    arXiv 2024

  7. [7]

    Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. 2021. Extracting training data from large language models. In 30th USENIX Security Symposium (USENIX Security 21), pages 2633--2650

    2021

  8. [8]

    Jiaao Chen and Diyi Yang. 2023. Unlearn what you want to forget: Efficient unlearning for llms. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12041--12052

    work page 2023

Show all 25 references
  1. [9]

    Minseok Choi, Daniel Rim, Dohyun Lee, and Jaegul Choo. 2024. https://arxiv.org/abs/2406.12329 Opt-out: Investigating entity-level unlearning for large language models via optimal transport . Preprint, arXiv:2406.12329

    2024 arXiv

  2. [10]

    Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, et al. 2020. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027

    2020 arXiv

  3. [11]

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representations (ICLR)

    2021

  4. [12]

    Joel Jang, Dongkeun Yoon, Sohee Yang, Sungmin Cha, Moontae Lee, Lajanugen Logeswaran, and Minjoon Seo. 2023. Knowledge unlearning for mitigating privacy risks in language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume...

    2023

  5. [13]

    Zhenghao Lin, Zhibin Gou, Yeyun Gong, Xiao Liu, Yelong Shen, Ruochen Xu, Chen Lin, Yujiu Yang, Jian Jiao, Nan Duan, et al. 2024. Rho-1: Not all tokens are what you need. arXiv preprint arXiv:2404.07965

    2024 arXiv

  6. [14]

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

    2022

  7. [15]

    Weitao Ma, Xiaocheng Feng, Weihong Zhong, Lei Huang, Yangfan Ye, Xiachong Feng, and Bing Qin. 2024. https://arxiv.org/abs/2406.15796 Unveiling entity-level unlearning for large language models: A comprehensive analysis . Preprint, arXiv:2406.15796

    2024 arXiv

  8. [16]

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

    2024 arXiv

  9. [17]

    Xander McCartney, Austin Young, and Dean Williamson. 2024. https://doi.org/10.36227/techrxiv.171863477.74163306/v1 Introducing anti-knowledge for selective unlearning in large language models

    2024

  10. [18]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36

    2024

  11. [19]

    Anil Ramakrishna, Yixin Wan, Xiaomeng Jin, Kai-Wei Chang, Zhiqi Bu, Bhanukiran Vinzamuri, Volkan Cevher, Mingyi Hong, and Rahul Gupta. 2025 a . Lume: Llm unlearning with multitask evaluations. arXiv preprint arXiv:2502.15097

    2025 arXiv

  12. [20]

    Anil Ramakrishna, Yixin Wan, Xiaomeng Jin, Kai-Wei Chang, Zhiqi Bu, Bhanukiran Vinzamuri, Volkan Cevher, Mingyi Hong, and Rahul Gupta. 2025 b . Semeval-2025 task 4: Unlearning sensitive content from large language models. arXiv preprint

    2025

  13. [21]

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

    2024 arXiv

  14. [22]

    Bichen Wang, Yuzhe Zi, Yixin Sun, Yanyan Zhao, and Bing Qin. 2024 a . Rkld: Reverse kl-divergence-based knowledge distillation for unlearning personal information in large language models. arXiv preprint arXiv:2406.01983

    2024 arXiv

  15. [23]

    Lingzhi Wang, Xingshan Zeng, Jinsong Guo, Kam-Fai Wong, and Georg Gottlob. 2024 b . https://arxiv.org/abs/2402.05813 Selective forgetting: Advancing machine unlearning techniques and evaluation in language models . Preprint, arXiv:2402.05813

    2024 arXiv

  16. [24]

    Yuanshun Yao, Xiaojun Xu, and Yang Liu. 2023. Large language model unlearning. arXiv preprint arXiv:2310.10683

    2023 arXiv

  17. [25]

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

    2024 arXiv

Pith tools

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