Pith. sign in

REVIEW 4 major objections 5 minor 35 references

UCD: Unlearning in LLMs via Contrastive Decoding

T0 review · 4 major / 5 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 Steering a model's output distribution at inference time can erase targeted knowledge from large language models without touching their weights.

desk verdict A clever, practical inference-time unlearning method with solid TOFU/MUSE News results when clean aux models exist, but the bootstrapping fallback undercuts itself and the theory is a restated assumption. read the letter →

arxiv 2506.12097 v1 pith:PADIBNRZ submitted 2025-06-12 cs.CL cs.CRcs.LGstat.ML

Vinith M. Suriyakumar , Ayush Sekhari , Ashia Wilson This is my paper
classification cs.CLcs.CRcs.LGstat.ML
keywords machineunlearningcontrastivedecodinginference-timelargelanguagemodelsTOFUbenchmarkMUSElogitsteeringLlama2-70B
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

This paper proposes that a large language model's unwanted knowledge can be erased at inference time by subtracting the logit difference between two small auxiliary models, one fine-tuned on the forget set and one on the retain set, from the large model's own logits. The authors claim this contrastive decoding update makes the model's outputs statistically indistinguishable from a model retrained without the forget data, while preserving or even improving utility on retained knowledge. They evaluate on TOFU and MUSE News and report that the method beats gradient-based and preference-optimization baselines on the forget-utility tradeoff, and that it scales to Llama2-70B using much smaller auxiliary models and fewer GPUs than fine-tuning baselines require.

What carries the argument

The update rule $\log P_{\text{aligned}}(y|x) \leftarrow \log P_{\text{corr}}(y|x) - \alpha\,(\log A_{\text{corr}}(y|x) - \log A_{\text{clean}}(y|x))$, which is a contrastive decoding step. The auxiliary difference $\Delta(y|x) = \log A_{\text{corr}}(y|x) - \log A_{\text{clean}}(y|x)$ is the mechanism: it is large and positive for tokens the forget-tuned model favors, so subtracting it suppresses those tokens in the reference model; a clipped variant, UCS, only suppresses and never boosts. The central object is the ratio between auxiliary and reference logit differences (equation 3), which the paper proves is sufficient for exact recovery of the clean distribution.

What would settle it

Train auxiliary models on a forget set where the proportionality in equation (3) can be measured directly on held-out prompts: if for a substantial fraction of tokens the sign of $\log A_{\text{corr}} - \log A_{\text{clean}}$ disagrees with the sign of $\log P_{\text{corr}} - \log P_{\text{clean}}$, then UCD's suppression direction is wrong and forget quality cannot approach the retrained model. Concretely, on TOFU Forget 10%, compute both differences over the forget QA pairs; a nonzero fraction of sign disagreements above a small threshold would falsify the proportionality premise that the proof and the effectiveness rely on.

Watch

Extended reading notes

Core claim

The central discovery is that unlearning need not modify weights at all. The paper shows that steering the next-token distribution of the reference model with the signed difference between a forget-tuned and a retain-tuned auxiliary model suppresses forget-set completions and boosts retain-set completions at generation time. In the idealized case where the auxiliaries equal the full models, the update recovers the retrained-from-scratch distribution exactly; under a proportionality assumption between auxiliary and full-model logit differences, the paper proves a single choice of the strength parameter yields the clean model. Empirically, UCD achieves forget quality that is indistinguishable from retraining on TOFU prompts while retaining higher utility, and it closely tracks the retrained model on MUSE News verbatim memorization, privacy leakage, and knowledge retention.

Load-bearing premise

The method needs a small 'clean' auxiliary model that was pretrained without the forget data, and the forget set must be separable enough that fine-tuning two auxiliaries captures the difference; when no clean model exists, such as for the Harry Potter corpus in MUSE Books, the paper does not apply its main claim and the task is excluded.

Editorial extensions

If this is right

  • If UCD is correct, any existing LLM can be unlearned without retraining or weight edits, by keeping two small fine-tuned auxiliaries at inference time.
  • The method's gray-box requirement, access to logits only and not gradients or parameters, means it applies to models served through APIs that expose logits.
  • Because the reference model is untouched, the same base model can be unlearned for different forget sets by swapping auxiliary pairs, making unlearning a per-request or per-tenant operation.
  • The computation shifts from training time to inference time, requiring three forward passes per token, which the paper shows is affordable at 70B scale with four GPUs.

Reading between the lines

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

  • A natural testable extension is to measure the proportionality in equation (3) directly on TOFU and MUSE: if the ratio bounds hold only within a narrow range of the strength parameter, the value chosen by tuning on a validation forget set may fail to transfer to new prompts from the same forget distribution.
  • The clean auxiliary requirement is the real-world bottleneck: for copyrighted corpora already present in every public pretraining corpus, no clean small model exists, and bootstrapping from approximate unlearned models inherits the baseline's own failure modes, as the paper itself shows when bootstrapping fails to help poor baselines like gradient ascent.
  • The method implicitly assumes tokenizer and vocabulary compatibility between reference and auxiliary models; the paper flags this limitation but does not test cross-tokenizer behavior, so a failure case could be probed by pairing models with different tokenizers.
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 / 5 minor

Summary. The paper proposes UCD, an inference-time unlearning method that steers a reference LLM's next-token distribution by subtracting a scaled version of the logit difference between two small auxiliary models, one fine-tuned on the forget set and one on the retain set; a clipped variant, UCS, is also introduced. Experiments on TOFU (5% and 10%) and MUSE News with Llama2-13B as the reference model and Llama2-7B auxiliary models report state-of-the-art forget-utility tradeoffs, forget quality approaching that of retraining, and successful scaling to Llama2-70B using Llama2-13B auxiliaries. The paper additionally proposes a bootstrapping procedure that replaces the clean auxiliary model with an existing unlearning baseline when no clean model is available. The main claims are that UCD substantially improves the forget-utility tradeoff on standard benchmarks and that it offers a practical, efficient route to unlearning in very large models.

Significance. If the empirical results hold, UCD is a valuable addition to the LLM unlearning toolbox: it performs unlearning at inference time, requires only logit access to the reference model, avoids gradient-based weight modification, and is demonstrated on models as large as Llama2-70B. The computational-cost comparison and the 70B experiments are concrete strengths, as are the ablation studies over sampling strategies and the α hyperparameter. The paper is also candid about several limitations. However, the theoretical justification rests on an unverified proportionality assumption, and the method's general applicability is sharply constrained by the need for a clean auxiliary model; the bootstrapping fallback is not validated by the reported numbers. These issues are load-bearing for the paper's broadest claims, though they do not necessarily invalidate the benchmark results obtained when exact clean auxiliary models are available.

major comments (4)
  1. [Section 3.1, Eqs. (3)-(5)] Proposition 1 is essentially tautological. The assumption in Eq. (3) states that the auxiliary logit difference is proportional to the reference model's corrupted-minus-clean logit difference; under that assumption, setting α = 1/m in Eq. (1) recovers P_clean by construction. No evidence is given that Eq. (3) holds for the Llama2-7B auxiliary models on TOFU or MUSE, and the constant m is never estimated from data. Instead, α is tuned per dataset over {0.01, 0.1, 0.5, 1.0} and the best value is reported (Section 5.1; Tables 12-13). The bounded-ratio extension in Eq. (5) has the same status. Consequently, the theoretical section does not predict or explain the observed gains; the empirical results carry the paper's weight on their own.
  2. [Section 5.2 and Table 11] The bootstrapping claim is not supported by the MUSE News results. Comparing NPO+RT with UCD to NPO+RT alone: VerbMem on D_forget worsens from 1.02 to 1.41, and KnowMem on D_retain drops from 34.27 to 28.09; only PrivLeak (64.58 to 63.91) and KnowMem on D_forget (28.78 to 25.53) improve. The text states that substituting an approximate clean model 'still provides benefits' while 'maintaining model utility,' but the utility metric actually deteriorates by about 6 points, and one of the two forget-quality metrics also worsens. Since this bootstrapping procedure is the only mechanism proposed for settings without a clean auxiliary model, the claim that UCD delivers state-of-the-art unlearning 'even without access to a clean model' is not established by the reported experiments.
  3. [Section 4 and Abstract] The evaluation is on TOFU (5% and 10%) and MUSE News, not MUSE as a whole. The paper states it was 'unable to obtain a clean model for the Books task' and therefore excludes MUSE Books. Books is precisely the setting where clean auxiliary models are unavailable, so excluding it makes the abstract's claim of evaluation on 'MUSE' and the contribution statement about 'standard machine unlearning benchmarks (TOFU, MUSE)' overly broad. The Discussion acknowledges the clean-model limitation, but the abstract and contributions should be revised to state that UCD is demonstrated only on TOFU and MUSE News and that applicability to settings such as MUSE Books remains open.
  4. [Section 3, opening paragraph, and Section 4] Essential training details of the auxiliary models are missing. The paper does not specify the fine-tuning procedure for A_corr and A_clean: number of epochs, learning rate, optimizer, batch size, LoRA versus full fine-tuning, data ordering, or how tokenizer mismatches between Llama2-7B and Llama2-13B/70B are handled, even though Section 7 identifies tokenizer matching as a requirement. Without these details, the main empirical results cannot be reproduced, and the claimed compute advantage is difficult to assess. A dedicated appendix with the full auxiliary-training protocol should be added.
minor comments (5)
  1. [Section 3.1] In the definition of P_clean, the phrase 'retraining-from-scratch without the retain set' should read 'without the forget set.'
  2. [Section 6.2] The UCS update is displayed as log P_aligned(y|x) ← log P_corr(y|x) − max{log A_corr(y|x) − log A_clean(y|x), 0}, but the surrounding text says 'where α > 0 is a hyperparameter'; α does not appear in the displayed formula, so either the formula should include α or the sentence should be corrected.
  3. [Sections 6.1-6.2] There are unresolved figure placeholders ('Figure ??' and 'Table ??') in the discussion of α sensitivity and in the UCD-versus-UCS comparison; these should be filled before publication.
  4. [Section 5.2 heading] The heading contains a typo: 'Bootsrapping' should be 'Bootstrapping.'
  5. [Table 10] The table reports minimum compute for only three methods and the text calls UCD 'optimal' in training and inference efficiency; either report the same compute requirements for all baselines or soften the wording to avoid an unsupported comparison.

Circularity Check

2 steps flagged · score 6.0 of 10

The theoretical 'why it works' in Section 3.1 builds the target clean-model distribution into its premises, so the formal recovery is a tautology; the benchmark comparisons are independent and are the real evidence.

  1. self definitional [Section 3.1, Eq. (2) (sanity check)]
    "First, as a sanity check, observe that if the auxiliary models are chosen to be the same size as the underlying models, that is, A clean = Pclean and A corr = Pcorr, and we set α = 1, then: log Pcorr(y|x)−α·(log A corr(y|x)−log A clean(y|x)) = log P clean(y|x).(2) In this special case, the contrastive differencing update in (1) exactly recovers the next token distribution corresponding to the model P clean that is retrained-from-scratch on the retain set."

    The equality is manufactured by the choice A_clean = P_clean and A_corr = P_corr: substituting these into (1) makes the auxiliary term cancel log P_corr and leave log P_clean by arithmetic. No property of the auxiliary models is established; the 'recovery' is just a renaming of the target distribution. Presenting this as illustrating how UCD enables unlearning assumes the very correspondence between auxiliary models and the target clean model that the method needs to establish.

  2. self definitional [Section 3.1, Proposition 1 and Eqs. (3)-(4)]
    "Proposition 1. Suppose that for any input prompt x, the auxiliary models A corr and A clean satisfy the relation: log Acorr(y|x)−log A clean(y|x)∝log P corr(y|x)−log P clean(y|x),(3) ... Then, there exists a choice of α that is independent of y such that the contrastive decoding procedure in (1) ensures that P aligned ≡P clean."

    Assumption (3) states that the auxiliary logit difference is proportional to the exact logit difference between the corrupted and clean models—the very quantity UCD aims to recover. From (4), setting α = 1/m turns (1) into an identity, so the proposition is the assumption restated under scaling. The subsequent bounded-ratio relaxation (5) is the same assumption with inequalities. Thus the theorem supplies no independent condition under which small auxiliary models track the target; it builds the target distribution into the premise.

full rationale

The empirical contribution of the paper is not circular: TOFU and MUSE results are judged by benchmark metrics that are independent of Section 3.1, and the comparison against retraining and baselines does not rely on Proposition 1 for its validity. The circularity is confined to the formal justification in Section 3.1, where both the sanity check (Eq. 2) and Proposition 1 assume access to, or proportionality with, the clean model P_clean that UCD is supposed to recover. This is a partial circularity: the theory does not predict UCD's success from independent first principles; it repackages the success condition as an assumption. The paper itself partially acknowledges the weakness ('we do not present this as a comprehensive explanation' and 'our approach currently lacks rigorous theoretical guarantees beyond the simplistic setting of Proposition 1'), which keeps the score below 8. Separately, the bootstrapping claim in Section 5.2 is not well supported by Table 11, where KnowMem on D_retain drops from 34.27 to 28.09 and VerbMem worsens; however, that is an evidence/consistency problem rather than a circular-derivation problem, so it does not add to the circularity score.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

No new physical or architectural entities are introduced; the method defines a virtual aligned distribution over logits rather than a new model. The main unexamined inputs are the proportional-difference assumption in Eq (3) and the existence of clean auxiliary training data.

free parameters (2)
  • alpha (UCD/UCS logit mixing weight) = 0.01, 0.1, 0.5, 1.0; per-task best selected (TOFU 5%: 0.5, TOFU 10%: 1.0, MUSE: 1.0)
    Controls how much contrastive signal is subtracted; results are reported for the best value per dataset after sweeping, so the headline tradeoff is partly a function of this tuned parameter.
  • proportionality constant m in Proposition 1 = unknown; set by Eq (4) when alpha=1/m
    Introduced in the theoretical section to make the auxiliary model difference equal the reference model difference; not estimated from data.
assumptions (5)
  • domain assumption There exists a public D_pretrain excluding D_forget and a base model A trained on it.
    Section 3 opening: 'suppose there exists some public dataset D_pretrain that does not contain D_forget, and a clean base model A'. Central to training A_corr and A_clean; authors acknowledge in Section 7 as a practical limitation.
  • ad hoc to paper Auxiliary logit difference is proportional to the reference model's corrupted-minus-clean logit difference (Eq 3).
    Section 3.1, Proposition 1. This assumption makes the derivation a tautology; it is not verified empirically.
  • domain assumption Forget and retain sets are sufficiently separable for targeted fine-tuning.
    Introduction says UCD applies 'when the data is sufficiently separable to allow targeted fine-tuning'; MUSE Books is excluded because no clean model was available.
  • domain assumption Tokenization schemes of reference and auxiliary models match.
    Section 7 limitations: 'requires careful matching of tokenization schemes between reference and auxiliary models; discrepancies could degrade quality.'
  • domain assumption Contrastive decoding improves text quality and diversity (Li et al. 2023, O'Brien and Lewis 2023).
    Used in Section 5.1 to attribute UCD's utility gain over retraining to CD's quality improvements; not re-tested here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UCD: Unlearning in LLMs via Contrastive Decoding." pith.science (2026). https://pith.science/paper/PADIBNRZ

@misc{pith2026250612097,
  author       = {Pith},
  title        = {Pith review of: UCD: Unlearning in LLMs via Contrastive Decoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PADIBNRZ}},
  note         = {Machine review of arXiv:2506.12097}
}
read the original abstract

Machine unlearning aims to remove specific information, e.g. sensitive or undesirable content, from large language models (LLMs) while preserving overall performance. We propose an inference-time unlearning algorithm that uses contrastive decoding, leveraging two auxiliary smaller models, one trained without the forget set and one trained with it, to guide the outputs of the original model using their difference during inference. Our strategy substantially improves the tradeoff between unlearning effectiveness and model utility. We evaluate our approach on two unlearning benchmarks, TOFU and MUSE. Results show notable gains in both forget quality and retained performance in comparison to prior approaches, suggesting that incorporating contrastive decoding can offer an efficient, practical avenue for unlearning concepts in large-scale models.

Figures

Figures reproduced from arXiv: 2506.12097 by the authors.

Figure 1
Figure 1. Illustration of contrastive decoding at inference time in UCD. In the top row, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Forget quality versus model utility averaged over three random seeds for TOFU [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Forget 5% 0.0 0.1 0.2 0.3 0.4 0.5 0.6 Model Utility 25 20 15 10 5 0 log(Forget Quality) NPO NPO + w/ UCD Retrain UCD [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (4 more)
Figure 6
Figure 6. Figure 6: Forget quality versus model utility TOFU 10% on Llama2-70B. UCD using Llama2- [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Forget 5% 0.0 0.1 0.2 0.3 0.4 0.5 0.6 Model Utility 25 20 15 10 5 0 Forget Quality (log) Greedy Nucleus (p = 0.9) Nucleus (p = 0.7) [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]
Figure 10
Figure 10. Figure 10: Forget 5% 0.0 0.1 0.2 0.3 0.4 0.5 0.6 Model Utility 30 25 20 15 10 5 0 Forget Quality (log) 0.01 0.1 0.5 1.0 [PITH_FULL_IMAGE:figures/full_fig_p026_10.png]
Figure 13
Figure 13. Figure 13: Forget 5% 0.0 0.1 0.2 0.3 0.4 0.5 0.6 Model Utility 35 30 25 20 15 10 5 0 log(Forget Quality) Grad Ascent Grad Diff NPO NPO + RT Retrain UCS UCD [PITH_FULL_IMAGE:figures/full_fig_p027_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 4 canonical work pages

  1. [3]

    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,

    arXiv

  2. [5]

    Regulation (eu) 2016/679 of the european parliament and of the council of 27 april

    European Union. Regulation (eu) 2016/679 of the european parliament and of the council of 27 april

    work page 2016

  3. [6]

    The curious case of neural text degeneration.arXiv preprint arXiv:1904.09751,

    Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration.arXiv preprint arXiv:1904.09751,

    arXiv 1904

  4. [9]

    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,

    arXiv

  5. [10]

    Soul: Unlocking the power of second-order optimization for llm unlearning.arXiv preprint arXiv:2404.18239,

    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,

    arXiv

  6. [12]

    Copyright violations and large language models.arXiv preprint arXiv:2310.13771,

    Antonia Karamolegkou, Jiaang Li, Li Zhou, and Anders Søgaard. Copyright violations and large language models.arXiv preprint arXiv:2310.13771,

    arXiv

  7. [13]

    Australia to consider european-style right to be forgotten privacy laws.The Guardian (Jan

    Pual Karp. Australia to consider european-style right to be forgotten privacy laws.The Guardian (Jan. 19, 2023).(Visited on 01/19/2023),

    work page 2023

  8. [14]

    Privacy adhering machine un-learning in nlp.arXiv preprint arXiv:2212.09573,

    Vinayshekhar Bannihatti Kumar, Rashmi Gangadharaiah, and Dan Roth. Privacy adhering machine un-learning in nlp.arXiv preprint arXiv:2212.09573,

    arXiv

Show all 35 references
  1. [15]

    Exact unlearning of finetuning data via model merging at scale

    Kevin Kuo, Amrith Setlur, Kartik Srinivas, Aditi Raghunathan, and Virginia Smith. Exact unlearning of finetuning data via model merging at scale. InICLR 2025 Workshop on Modularity for Collaborative, Decentralized, and Continual Deep Learning. Meghdad Kurmanji, Peter Triantafi...

    2025

  2. [16]

    Faster machine unlearning via natural gradient descent.arXiv preprint arXiv:2407.08169,

    Omri Lev and Ashia Wilson. Faster machine unlearning via natural gradient descent.arXiv preprint arXiv:2407.08169,

    arXiv

  3. [17]

    The wmdp benchmark: Measuring and reducing malicious use with unlearning.arXiv preprint arXiv:2403.03218,

    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,

    arXiv

  4. [18]

    URL https://aclanthology.org/2023.acl-long.687/

    doi: 10.18653/v1/2023.acl-long.687. URL https://aclanthology.org/2023.acl-long.687/. Bo Liu, Qiang Liu, and Peter Stone. Continual learning and private unlearning. InConference on Lifelong Learning Agents, pp. 243–254. PMLR,

    2023 doi

  5. [19]

    Large language model unlearning via embedding-corrupted prompts.arXiv preprint arXiv:2406.07933, 2024a

    Chris Yuhao Liu, Yaxuan Wang, Jeffrey Flanigan, and Yang Liu. Large language model unlearning via embedding-corrupted prompts.arXiv preprint arXiv:2406.07933, 2024a. 14 Sijia Liu, Yuanshun Yao, Jinghan Jia, Stephen Casper, Nathalie Baracaldo, Peter Hase, Xiaojun Xu, Yuguang Ya...

    arXiv

  6. [20]

    An adversarial perspective on machine unlearning for ai safety.arXiv preprint arXiv:2409.18025,

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

    arXiv

  7. [21]

    Locating and editing factual associations in gpt.arXiv preprint arXiv:2202.05262, 2022a

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt.arXiv preprint arXiv:2202.05262, 2022a. Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. Mass- editing memory in a transformer.arXiv prepri...

    arXiv

  8. [22]

    Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. Memory-based model editing at scale.arXiv preprint arXiv:2110.11309, 2022a. Eric Mitchell, Charles Lin, Antoine Bosselut, Christopher D Manning, and Chelsea Finn. Memory-based model editing...

    arXiv

  9. [23]

    A survey of machine unlearning.arXiv preprint arXiv:2209.02299,

    Thanh Tam Nguyen, Thanh Trung Huynh, Phi Le Nguyen, Alan Wee-Chung Liew, Hongzhi Yin, and Quoc Viet Hung Nguyen. A survey of machine unlearning.arXiv preprint arXiv:2209.02299,

    arXiv

  10. [24]

    Contrastive decoding improves reasoning in large language models.arXiv preprint arXiv:2309.09117,

    Sean O’Brien and Mike Lewis. Contrastive decoding improves reasoning in large language models.arXiv preprint arXiv:2309.09117,

    arXiv

  11. [25]

    URL https://arxiv.org/abs/2309. 09117. Martin Pawelczyk, Seth Neel, and Himabindu Lakkaraju. In-context unlearning: Language models as few shot unlearners.arXiv preprint arXiv:2310.07579,

    arXiv

  12. [26]

    Rethinking llm memorization through the lens of adversarial compression.arXiv preprint arXiv:2404.15146,

    Avi Schwarzschild, Zhili Feng, Pratyush Maini, Zachary C Lipton, and J Zico Kolter. Rethinking llm memorization through the lens of adversarial compression.arXiv preprint arXiv:2404.15146,

    arXiv

  13. [27]

    Model evaluation for extreme risks.arXiv preprint arXiv:2305.15324,

    Toby Shevlane, Sebastian Farquhar, Ben Garfinkel, Mary Phuong, Jess Whittlestone, Jade Leung, Daniel Kokotajlo, Nahema Marchal, Markus Anderljung, Noam Kolt, et al. Model evaluation for extreme risks.arXiv preprint arXiv:2305.15324,

    arXiv

  14. [28]

    Detecting pretraining data from large language models

    Weijia Shi, Anirudh Ajith, Mengzhou Xia, Yangsibo Huang, Daogao Liu, Terra Blevins, Danqi Chen, and Luke Zettlemoyer. Detecting pretraining data from large language models. arXiv preprint arXiv:2310.16789,

    arXiv

  15. [29]

    Muse: Machine unlearning six-way evaluation for language models.arXiv preprint arXiv:2407.06460,

    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,

    arXiv

  16. [30]

    Guardrail baselines for unlearning in llms.arXiv preprint arXiv:2403.03329,

    Pratiksha Thaker, Yash Maurya, and Virginia Smith. Guardrail baselines for unlearning in llms.arXiv preprint arXiv:2403.03329,

    arXiv

  17. [32]

    Genarm: Reward guided generation with autoregressive reward model for test-time alignment.arXiv preprint arXiv:2410.08193,

    Yuancheng Xu, Udari Madhushani Sehwag, Alec Koppel, Sicheng Zhu, Bang An, Furong Huang, and Sumitra Ganesh. Genarm: Reward guided generation with autoregressive reward model for test-time alignment.arXiv preprint arXiv:2410.08193,

    arXiv

  18. [33]

    Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities.arXiv preprint arXiv:2408.07666,

    Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities.arXiv preprint arXiv:2408.07666,

    arXiv

  19. [34]

    Large language model unlearning.arXiv preprint arXiv:2310.10683, 2023a

    Yuanshun Yao, Xiaojun Xu, and Yang Liu. Large language model unlearning.arXiv preprint arXiv:2310.10683, 2023a. Yunzhi Yao, Peng Wang, Bozhong Tian, Siyuan Cheng, Zhoubo Li, Shumin Deng, Huajun Chen, and Ningyu Zhang. Editing large language models: Problems, methods, and oppor...

    arXiv

  20. [35]

    focuses on updating specific behaviors or outputs of trained models, typically targeting localized knowledge adjustments or corrections of individual predictions. It typically involves precise interventions at the token or representation level, enabling targeted updates withou...

    2022

  21. [2017]

    Large scale knowledge washing.arXiv preprint arXiv:2405.16720,

    Yu Wang, Ruihan Wu, Zexue He, Xiusi Chen, and Julian McAuley. Large scale knowledge washing.arXiv preprint arXiv:2405.16720,

    arXiv

  22. [2018]

    Open problems in machine unlearning for ai safety.arXiv preprint arXiv:2501.04952,

    Fazl Barez, Tingchen Fu, Ameya Prabhu, Stephen Casper, Amartya Sanyal, Adel Bibi, Aidan O’Gara, Robert Kirk, Ben Bucknall, Tim Fist, et al. Open problems in machine unlearning for ai safety.arXiv preprint arXiv:2501.04952,

    arXiv

  23. [2019]

    Editing models with task arithmetic

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. 13 arXiv preprint arXiv:2212.04089,

    arXiv

  24. [2021]

    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, pp. 463–480. IEEE,

    2015

  25. [2022]

    Knowledge sanitization of large language models

    Yoichi Ishibashi and Hidetoshi Shimodaira. Knowledge sanitization of large language models. arXiv preprint arXiv:2309.11852,

    arXiv

  26. [2023]

    Towards scalable exact machine unlearning using parameter- efficient fine-tuning.arXiv preprint arXiv:2406.16257,

    Somnath Basu Roy Chowdhury, Krzysztof Choromanski, Arijit Sehanobish, Avinava Dubey, and Snigdha Chaturvedi. Towards scalable exact machine unlearning using parameter- efficient fine-tuning.arXiv preprint arXiv:2406.16257,

    arXiv

  27. [2024]

    Split, unlearn, merge: Leveraging data attributes for more effective unlearning in llms

    Swanand Ravindra Kadhe, Farhan Ahmed, Dennis Wei, Nathalie Baracaldo, and Inkit Padhi. Split, unlearn, merge: Leveraging data attributes for more effective unlearning in llms. arXiv preprint arXiv:2406.11780,

    arXiv

Pith tools

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