acceptodds
Under review as a conference paper at ICLR 2027

Nonmember-Law Matching for Suppressing Unlearning Signatures

Abstract

Machine unlearning can suppress a language model's knowledge of forget data while leaving it vulnerable to membership inference: forget examples remain distinguishable from data the model was never trained on, revealing which records were used in training or targeted for removal. Existing LLM unlearning objectives lower forget likelihoods or redirect internal representations without reference to nonmember behavior, so unlearning can stop short or overshoot, and both leave identifiable signatures. We propose Nonmember-Law Matching (NML), which uses the original model's likelihoods on nonmembers as the target for forget examples. NML matches the distribution of forget-example negative log-likelihoods to this nonmember reference and preserves retained behavior through distillation, without requiring a retrained model. On two unlearning benchmarks, TOFU and MUSE-News, NML substantially reduces unlearning signatures under multiple likelihood-based attacks while maintaining forgetting and utility comparable to or better than strong baselines, and question-only identification attacks provide further evidence of this reduction. These results suggest that calibrating against nonmember behavior is a simple and effective way to reduce the identifiable traces of unlearning.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
machine unlearning, AI security, LLM

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_BeojOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Nonmember-Law matching (NLM) as an algorithm for performing machine unlearning in large language models guided by the principle that an unlearned model should treat the forget set no different than a held out test set of the same distribution that was not trained on. The algorithm achieves unlearning by balancing score-matching inspired optimization using a reference set and the forget set alongside distilling the retain set from the original model to preserve model utility. The authors evaluate their method on two unlearning benchmarks, and achieve state-of-the-art performance on a variety of metrics over a number of baselines.

Overall this paper presents a novel and interesting approach to unlearning, which is sufficiently justified. The work is well written and well presented, but lacks in substantial areas and raises nontrivial questions. Given sufficient clarification to the points outlined below, I would be very open to increasing my score.

Critical strengths
  • The work positions itself well within the larger literature and has comprehensive evaluation against baselines which are clearly outlined and well presented. NML seems to outperform most state-of-the-art methods for LLM unlearning through its novel procedure.
  • The unlearning method in question is intuitive and interesting, and the energy-based score-matching approach is a powerful technique to leverage for unlearning.
Critical weaknesses
  • While I am happy that the authors have provided code alongside their submission - the code does not include evaluation scripts or scripts for reference data generation. The second in particular is in my opinion a critical part of the work.
  • All evaluation experiments are performed on the Llama family of models, which calls into question how well NML works on other model families. Can the authors provide such an experiment?
  • The method relies greatly on the prerequisite that the distribution of the reference set is approximately equivalent to the distribution of the forget set. This is only outlined in Section 2, and the presentation would be improved if this was stated in the introduction.
  • It seems unlikely to me that in practical settings a held-out reference set of the same distribution as the forget set is available, and therefore I argue that the reference set construction case is a far more interesting exploration than when a held-out set is available. The performance of NML on the reference-set construction case is lacking sufficient empirical evidence.
  • The authors note that the Retrained model does not obtain zero gaps indicating that forget and held-out records remain distinguishable. This calls into question the validity of the reference set, and the target of the optimization performed in NML. Can this be chalked up to randomness, in which case can the authors report retraining across seeds, or does this suggest that reference sets must be more carefully chosen or constructed. I would welcome discussion from the authors on this point.
  • A milder point is that the title is misleading since this technique is tested exclusively for LLMs.
Questions for authors
  • In Figure 1, NPO and SimNPO cross the 0.5 threshold and go below it while NML remains close to 0.5 the entire time. This makes me ask, had the baselines achieved unlearning when 0.5 was reached?
  • When the reference set is not of the same distribution as the forget set, how does the method behave? It seems to me as if it is hard to verify if the distributions are the same if that is not known apriori.
8 comments
AuthorsOpenReview ↗

Response to Reviewer Beoj (1/3): W1–W3 & W6 — code, model family, introduction, title

Thank you for the careful review.

W1: code for reference generation and evaluation

Since you identify reference generation as the critical part, we prepared an anonymous repository for it: https://anonymous.4open.science/r/ref-construction-release-C364/

For each benchmark, it includes the constructor script, inputs, prompts, generation settings, acceptance rules and thresholds, three constructed references, and reports listing rejected generations. Neither constructor reads the benchmark's held-out split or the retrained model, and examples are accepted by fixed rules on their text, never by their score under the model to be unlearned. We regenerated all three TOFU references and reran training and evaluation. The resulting means differ from Table 4 by at most 0.004 at the table's precision. For evaluation, we follow the protocol of the OpenUnlearning benchmark and use its public implementations of the metrics and attacks (Sec. 3.1); what we add is the split of held-out nonmembers into a reference fold and an attack fold, described in the same section. We will add the prompts to Appendix A and release the training and evaluation code with the final version.

W2: another model family

NML also works outside the Llama family. We added Phi-1.5 on TOFU Forget10 (3 seeds, same setting as Table 1):

MethodKS-DMULosszlibMin-K%Min-K%++
Full0.3450.3950.4980.4990.4980.499
Retrain00.4150.1360.2040.1100.006
NPO0.1770.4020.2070.2110.0930.323
SimNPO0.2150.3980.2440.1420.2290.225
UnDIAL0.2600.3820.4510.3580.4870.500
NML0.1480.4020.0410.0690.0380.167

NML has the lowest KS-D and the lowest gap under each attack among the unlearning methods shown, with the same MU as NPO.

We also ran the larger Phi-3.5-mini (3.8B) with the hyperparameters of the Llama setting unchanged (3 seeds). NML has the lowest gap under three of the four attacks (Loss 0.099, zlib 0.031 and Min-K%++ 0.013, against 0.118, 0.152 and 0.100 for NPO, the best baseline there); on Min-K%, NPO is slightly lower (0.079 against 0.089). We will add both tables to the appendix. On a third family, Gemma-7B-it, we have first results under the same unchanged hyperparameters. NML reaches Loss, zlib and Min-K% gaps of 0.045, 0.157 and 0.034 (3 seeds), at or below Retrain's (0.070, 0.162, 0.080), whereas NPO overshoots (Loss gap 0.41–0.49 in the two seeds completed so far). Given the time available, the remaining baselines on Gemma and further model families are still running; we will add them to the appendix.

W3 & W6: the prerequisite in the introduction; the title

We agree with both. The third paragraph of the introduction will state the prerequisite directly after the sentence on where the reference comes from:

"NML assumes that the reference is comparable to the forget set in task, example format and scoring unit, and that unlearning leaves the scores of nonmembers approximately unchanged."

If the venue permits a title change, the title will become "Nonmember-Law Matching for Suppressing Unlearning Signatures in Language Models"; in either case, the abstract and introduction will state that the method and its evaluation are for language models.

AuthorsOpenReview ↗

Response to Reviewer Beoj (2/3): W4 & W5 — constructed references; Retrain's nonzero gaps

W4: evidence for constructed references

We agree that the constructed case is important in practice. We add evidence in three parts: constructed references on a second, larger model (TOFU); the LLM-constructed news references, which remain weaker; and, beyond construction, public human-written text as another source of reference.

Constructed references on two models (TOFU). The table compares NML trained with the held-out reference and with a constructed one (3 seeds per row):

ModelReference used by NMLLoss gapMin-K% gapMU
Llama-3.2-1BHeld-out (Table 4)0.0520.0440.600
Llama-3.2-1BConstructed, references 1–3 (Table 4)0.077–0.1100.069–0.1020.600–0.601
Llama-2-7BHeld-out (Table 1)0.0490.2360.631
Llama-2-7BConstructed, reference 1 (new)0.0550.1850.632

On both models, NML with a constructed reference stays close to NML with the held-out reference, at the same utility, and far below SimNPO's loss gap (0.237 on 1B and 0.410 on 7B, Table 1). We will add the 7B row to Table 4.

LLM-constructed news (MUSE-News). These references are coarser, because a language model has difficulty generating long text that is close enough to human-written news. NML's loss gap with them (0.263–0.331, Appendix B) is still below those of NPO (0.498), UnDIAL (0.498) and SimNPO (0.455). Our answer to Q2 gives a check that identifies such a reference before unlearning.

Beyond construction: public text (MUSE-News). This is not a constructed reference, and it applies only when public text of the same genre exists. With BBC articles (the source of MUSE-News) published after the benchmark's period and filtered for overlap with MUSE (3 seeds), NML reaches a loss gap of 0.052, the same as Retrain, against 0.123 with the held-out reference (Table 2).

Taken together, constructed references are supported on TOFU at two model scales and remain weaker for long-form news; where public text of the same genre exists, it is an alternative.

W5: Retrain's nonzero gaps, and what they mean for the reference and the target

It is not randomness. We retrained five models with different seeds on TOFU 1B. For the loss attack, the AUC is 0.396 ± 0.001 across these models and 0.388 for the official Retrain of Table 1, corresponding to gaps of 0.104 and 0.112. The gap reflects a systematic difference between the forget and held-out authors.

What this means for the target. NML's target, the original model's NLLs on the reference, can differ from the ideal target, Retrain's NLLs on the forget set, for two reasons. The first is the model: the original model has seen the forget set. This effect is small: on the same held-out examples, the original model and Retrain are at a KS distance of 0.030–0.045 on TOFU and 0.047–0.055 on MUSE-News. The second is the data: under Retrain, the held-out and forget sets are at 0.215–0.245 on TOFU and 0.133–0.141 on MUSE-News. Retrain's nonzero gaps show this second difference. NML moves the forget NLLs to the reference, so they differ from Retrain's by about this amount, and its gap against held-out examples can fall below Retrain's; we do not read this as better than Retrain. The difference shrinks as the reference becomes more comparable to the forget set, and two results indicate that it is small enough in practice: NML has the lowest KS-D to Retrain on TOFU 1B (0.061, Table 1), and references that do not come from the held-out distribution (constructed QA, public news) still give low gaps against held-out examples.

How much the choice of reference matters. References within a KS distance of 0.30 of the ideal target (Sec. 4.1) give loss gaps of 0.049–0.123 after NML, in no particular order of distance. The LLM-constructed news, at 0.375–0.422, gives 0.263–0.331. The distance thus separates a clearly mismatched reference from acceptable ones and does not rank them. Our answer to Q2 gives a check that needs no Retrain.

AuthorsOpenReview ↗

Response to Reviewer Beoj (3/3): Q1 & Q2 — Figure 1, mismatched references

Q1: had the baselines achieved unlearning when the AUC reached 0.5?

Not fully: at the crossing, forgetting is not yet complete. We tracked Min-K% AUC and KS-D around the crossings in Figure 1 (TOFU 1B, 3 seeds).

StepSimNPO: KS-D / AUCNPO: KS-D / AUCNML: KS-D / AUC
600.133 / 0.5080.193 / 0.5970.103 / 0.557
720.097 / 0.3570.180 / 0.3900.088 / 0.544
840.088 / 0.2960.168 / 0.2810.090 / 0.578

SimNPO is closest to chance at step 60 (AUC 0.508), where its KS-D is 0.133. Its KS-D reaches 0.088 at step 84, where its AUC has fallen to 0.296, and 0.076 at the final step (Table 1). NPO crosses between steps 60 and 72 with a KS-D of 0.180–0.193, against 0.151 at the final step. For both baselines, the step with a near-chance AUC and the step with the best agreement with Retrain are different steps. The crossing is also brief: within 12 updates, SimNPO's AUC falls from 0.508 to 0.357 and NPO's from 0.597 to 0.390. NML's AUC stays within 0.544–0.578 over the same steps, and its KS-D at the final step is 0.061 (Table 1).

Q2: behavior under a mismatched reference, and how to verify a reference

Behavior. NML degrades gradually, not abruptly. On TOFU 1B we replaced a growing share of the held-out reference with QA about real well-known authors (different difficulty) or about fictional companies (different domain), over 3 seeds:

Share replacedReal authors: Loss / Min-K% gapFictional companies: Loss / Min-K% gap
0%0.052 / 0.0440.052 / 0.044
20%0.079 / 0.0750.083 / 0.074
50%0.120 / 0.1140.115 / 0.109
100%0.255 / 0.2500.216 / 0.211

MU stays at 0.598–0.601. At 50% replacement, both gaps remain well below SimNPO's (0.237 / 0.252). Full replacement brings them close to that level.

Verification. A mismatch can be detected before unlearning, without a retrained model. We score the forget set and the candidate reference with a model that was trained on neither, and compute the KS distance between the two NLL distributions. On MUSE-News, with three scoring models (the checkpoint before fine-tuning, Phi-1.5 and Llama-3.2-1B), the distance is 0.05–0.14 for public news, 0.09–0.27 for the held-out set and 0.98–1.00 for LLM-constructed news, which is also the reference with the largest final loss gap (0.263–0.331, against 0.052 and 0.123). On TOFU, under three scoring models, the distance is 0.30–0.32 for the held-out reference, 0.43–0.59 when the reference is fully replaced by QA about real authors, and 0.57–0.65 when it is replaced by QA about fictional companies; the loss gaps after NML are 0.052, 0.255 and 0.216. The distance is read relative to a comparable reference of the same task, not against a fixed threshold. We regard the check as a screening test: it separates mismatched references from comparable ones, and we do not claim that it predicts the final loss gap among comparable references. We will add these results to a new appendix on reference validity.

Reviewer_9exeOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the problem that when unlearning forget set, existing unlearning methods can stop short or overshoot, rendering forget data identifiable under membership inference attacks (MIAs). To address this privacy concern, the paper proposes Nonmember-Law Matching (NML). The key idea is the forget data's log likelihood distribution should be trained to match the corresponding non-members distribution (instead of matching the mean or minimised). To preserve model utility, the unlearned model also minimise the KL divergence of the token distribution on the retain set. The empirical results show that NML successfully reduce the success of MIAs under various sequence and token level attack signatures.

Critical strengths
  1. The paper identifies an important problem of improving privacy protection (decreasing MIA success rate) after unlearning and proposes an empirically effective solution. Different attack signatures are considered.
  2. The perspective on matching the loss distribution instead of matching/minimizing the mean is novel.
Critical weaknesses
  1. To demonstrate the superiority of NML, the paper should provide a more comprehensive comparison with existing baseline methods.
    • Fig 1 shows that NPO/SimNPO has overshoot unlearning. In practice, it may be possible to track the AUC and stop earlier.
    • Table 1 only reports the unlearned models after a fixed number of iterations/hyperparameter setting.
    • To be fair, NML should be compared with more unlearning methods that also train on the retain set, such as gradient difference from the TOFU paper. There are more baseline methods in OpenUnlearning.
  2. The paper does not empirically analyse when/why NML is likely to perform better. Only one forget/retain split on TOFU/MUSE is used.
Questions for authors
  1. Discuss or add additional experiment results to address weakness 1.
  2. When is NML less effective? Will it work well on a non QA dataset? For example, on other datasets or forget/retain split, the forget and retain datasets may be more similar. This makes it harder to control the forget distribution of NLL without simultaneously affecting retain utility/NLLs.
  3. Sec 5.3 provides empirical comparisons of the computation cost. Can the authors directly discuss the computation cost/complexity of NML (e.g., distribution matching) as compared to baselines?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 9exe (1/2): W1 & Q1 — early stopping, settings and more baselines

Thank you for the review and the concrete suggestions.

W1(a) & Q1: can the baselines be stopped earlier?

Stopping near the crossing brings the Min-K% AUC close to 0.5, but forgetting is not yet complete there, and the stopping point is hard to hit. We tracked Min-K% AUC and KS-D around the crossings in Figure 1 (TOFU 1B, 3 seeds).

StepSimNPO: KS-D / AUCNPO: KS-D / AUCNML: KS-D / AUC
600.133 / 0.5080.193 / 0.5970.103 / 0.557
720.097 / 0.3570.180 / 0.3900.088 / 0.544
840.088 / 0.2960.168 / 0.2810.090 / 0.578

Forgetting is incomplete at the crossing. SimNPO is closest to chance at step 60 (AUC 0.508), where its KS-D is 0.133. Its KS-D reaches 0.088 at step 84, where its AUC has fallen to 0.296, and 0.076 at the final step (Table 1). NPO crosses between steps 60 and 72 with a KS-D of 0.180–0.193, against 0.151 at the final step. For both baselines, the step with a near-chance AUC and the step with the best agreement with Retrain are different steps.

The crossing is brief. Within 12 updates, SimNPO's AUC falls from 0.508 to 0.357 and NPO's from 0.597 to 0.390, so shifting the stopping step by a few updates moves the AUC by 0.15–0.21. NML needs no stopping rule: its AUC stays within 0.544–0.578 over the same steps, and its KS-D at the final step is 0.061 (Table 1). At the final step, the Min-K% gaps are 0.252 for SimNPO, 0.321 for NPO and 0.044 for NML (Table 1).

W1(b): one number of iterations and one hyperparameter setting

We checked both. The table above covers the number of iterations. For hyperparameters, OpenUnlearning releases checkpoints of eight methods on this task, each trained with its own recipe over a grid of 40–54 settings (398 checkpoints). We evaluated all of them under the attacks of Table 1. We summarize each checkpoint by its worst-case attack gap, the largest of its Loss, zlib, Min-K% and Min-K%++ gaps. For each method, the table gives the released setting with the lowest worst-case gap, that is, the setting most favorable to that method, next to NML at its single default setting:

Method (number of released settings)Worst-case attack gapKS-DMU
IdkDPO (54)0.0700.0950.540
RMU (54)0.0900.0880.249
NPO (54)0.0960.1630.172
AltPO (54)0.1030.1350.366
SimNPO (48)0.1120.1130.554
GradDiff (40)0.1200.2280.589
UnDIAL (54)0.2350.2830.122
IdkNLL (40)0.4730.2800.509
NML (1, default; 3 seeds, Table 5)0.1140.0630.600

Five of these settings match or improve on NML's worst-case gap. Each has higher KS-D and lower MU. Across all released checkpoints, none matches or improves on both NML's worst-case gap and KS-D. NML itself is stable: in 10 of 11 settings of λ_R, λ_F, learning rate, microbatch size and LoRA rank, the loss and Min-K% gaps stay below 0.12 (SimNPO: 0.237 and 0.252). We will add both analyses to the appendix.

W1(c): more baselines that train on the retain set

We added GradDiff from OpenUnlearning's released checkpoints for the same task and model. Of its 40 released settings we report the one with the lowest worst-case attack gap: 0.120, with KS-D 0.228 and MU 0.589. Its attack gap and utility are close to NML's (0.114 and 0.600), and its KS-D is higher (0.228 against 0.063). We will add it to Table 1, marked as using its own recipe, as RMU and LUNAR are. The other OpenUnlearning methods (RMU, AltPO, IdkDPO, IdkNLL) are included in the table under W1(b). These comparisons support the joint result in Table 1: NML reduces attack gaps while maintaining agreement with Retrain and retained utility.

AuthorsOpenReview ↗

Response to Reviewer 9exe (2/2): W2, Q2 & Q3 — when NML works, and its cost

W2 & Q2: when and why NML performs better, and when it is less effective

Why NML performs better. The baselines' forget terms give a direction but no endpoint, so their attack AUC depends on where training stops; NML's forget term has an endpoint, the nonmember NLL distribution. In W1(a), the baselines cross 0.5 and move farther from chance by step 84, while NML stays between 0.544 and 0.578 over the three shown checkpoints.

More than one split. We added TOFU Forget05 and Forget01 (1B, 3 seeds, final step):

SplitNML: Loss / Min-K% gapSimNPONPO
Forget050.037 / 0.0270.260 / 0.2780.388 / 0.373
Forget010.054 / 0.0430.308 / 0.3560.423 / 0.443

NML has lower loss and Min-K% gaps on both splits, with MU within 0.003 of the listed baselines. We will add both, with all baselines, to the appendix.

Non-QA data. NML also works on documents: MUSE-News is document unlearning on 2048-token chunks, and NML has the lowest loss and zlib gaps among the unlearning methods in Table 2.

When NML is less effective. As you anticipate, the limit is reached when the forget set, the retain set and the nonmembers share content. MUSE-Books is such a case: all three come from one book series. There NML still removes verbatim memorization (VerbMem-F from 0.997 to 0.108; Retrain 0.144), at a cost in retained knowledge (KM-R from 0.693 to 0.593), and the privacy metric does not improve. The cause is specific to this structure: unlearning some volumes also changes the model's NLL on the other volumes (median holdout NLL from 4.06 to 4.36), so a target fixed before unlearning no longer describes the nonmembers. When the nonmembers share no content with the forget set, this shift is negligible (News: mean held-out NLL from 1.90 to 1.95), which is the case for different authors in TOFU and different articles in News. We will state this scope in the limitations (Sec. 6).

Q3: computational cost of distribution matching

The cost of NML is that of the shared-LoRA baselines plus a negligible term. Per update it runs the same model passes as those baselines: forward and backward on B forget and B retain examples, and a forward pass of the frozen original model on the retain examples for the distillation term. The matching term adds bn + b² absolute differences of scalars per microbatch, that is O(B(n + b)) scalar operations per update, and no model pass. The one-off cost is n forward passes without gradients to score the reference. This agrees with the measurements of Sec. 5.3: 3.84–3.85 minutes for NML against 3.80–3.84 for SimNPO, with the same peak memory. We will add this analysis to Sec. 2.3.

Thank you for the detailed responses which have addressed my concerns. The additional experiments for other reviews were also useful. I will keep my positive rating of the work.

Can the authors update Sec. 6 with the new limitations and scope of the work?

AuthorsOpenReview ↗

Response to Reviewer 9exe's follow-up

Thank you for the follow-up and for your positive assessment. We are glad that the responses addressed your concerns.

Yes, we will update Sec. 6 with the limitations and scope identified in this discussion. Since the discussion with the other reviewers is still ongoing, we will make this update once it has concluded, so that the revised section reflects all of the points raised.

Reviewer_ueW4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper starts from a problem in LLM unlearning: after unlearning, forget examples may still look different from examples the model was never trained on, allowing an attacker to identify them. Simply lowering forget-example likelihood can overshoot, leaving the likelihood unusually low and potentially revealing that an example was deliberately unlearned.

The paper focuses on reducing unlearning signatures under multiple likelihood-based attacks while maintaining forgetting performance and retained utility. It shows that previous methods such as UnDIAL and NPO do not consistently achieve an AUC of 0.5 under the Min-K% attack and notes that their objectives do not explicitly refer to nonmember behavior.

The proposed Nonmember-Law Matching uses the original model’s negative log-likelihoods on nonmembers as a reference target for forget examples. It trains the model to align the distribution of scalar NLL scores on forget examples matches this reference distribution using energy distance, with a distillation term on retained data to preserve utility.

The reference comes from held-out benchmark data or examples constructed following the task format. The paper conducts experiments on TOFU with 1B and 7B models and on MUSE-News with a 7B model, measuring forgetting, retained utility, four likelihood-based attacks, and question-only identification.

Critical strengths
  1. This paper proposes a direct, attacker-free objective for nonmember-calibrated LLM unlearning through a cached NLL reference and energy-distance matching. Rather than continuously suppressing forget likelihoods or matching only a mean score, it adjusts the distribution of scalar forget-example NLLs toward that reference.
  2. The paper is clearly written and organized in a way that is easy to follow. The motivation and proposed method are naturally connected. The discussion of limitations helps clarify the scope of the claims.
Critical weaknesses
  1. NML uses the original model, which has been trained on the forget data, to construct the reference. Its behavior on nonmembers is a proxy for how a model trained without the forget set would score forget examples, but these distributions need not coincide. Being a nonmember does not, by itself, make an example a suitable reference for the forget set; record difficulty and domain can affect NLL independently of membership. Although the paper investigates reference comparability through direct comparisons with Retrain on TOFU and news, the evidence remains limited, particularly given the weaker results with constructed news references. These references appear comparable at the document level but are considerably less comparable at the 2048-token chunk level used for training.

  2. Nonmember calibration and membership-guided unlearning have precedents, including SCRUB+R [1] and the membership-guided approaches discussed in the paper. NML’s particular combination of a cached NLL reference, energy-distance matching, and retain distillation may provide a useful formulation, but its contribution should be distinguished more clearly from these earlier principles. SCRUB+R is not discussed, while the comparison with earlier membership-guided approaches could be developed further. Although energy matching improves the reported forgetting metric, stronger comparisons are needed to establish the additional benefit of the proposed distribution-matching construction.

[1] Kurmanji et al. Towards Unbounded Machine Unlearning.

  1. The paper conducts experiments only on TOFU Forget10 and MUSE-News, without the Books domain of MUSE. The main MUSE comparison reports knowledge memorization, retained utility, and four likelihood-based attack gaps, but does not include verbatim memorization, which appears only in the supplementary reference-construction study. It also does not report the original MUSE PrivLeak metric, instead measuring absolute AUC deviations from 0.5 rather than deviations from the retrained reference. Broader domain coverage and a more complete evaluation would help establish whether signature suppression accompanies reliable content removal and remains effective beyond the reported settings.
Questions for authors
  1. The evaluation could be strengthened by including MUSE-Books to assess whether the benefits extend beyond TOFU and news, particularly given the method’s dependence on reference comparability. WMDP could also provide complementary evidence of applicability to broader knowledge-removal settings. Where feasible, reporting verbatim memorization across the main baselines and the original MUSE PrivLeak metric would help provide a more complete picture of unlearning performance.

  2. Could the authors provide further justification for using the original model’s nonmember NLL distribution as a reference for forget examples? In particular, how can reference suitability be assessed without access to Retrain, given possible differences in difficulty, domain, and scoring unit? Further discussion of the document-level versus chunk-level discrepancy in the constructed-news experiments would help clarify when this approximation is reliable.

  3. Could the authors further clarify the contribution beyond earlier nonmember-calibrated and membership-guided approaches? Since mean matching preserves most of the reported attack-gap improvements, additional justification or controlled comparisons would help establish what energy-distance matching contributes beyond selecting a nonmember target.

7 comments
AuthorsOpenReview ↗

Response to Reviewer ueW4 (1/3): W1 & Q2 — is the nonmember reference a valid target?

Thank you for the careful review.

W1 & Q2: is the original model's nonmember NLL distribution a valid reference?

(1) Justification. NML uses the original (Full) model's NLLs on nonmembers as a proxy for the ideal target, Retrain's NLLs on the forget set. The proxy is valid when the reference is comparable to the forget set: the original model's exposure to the forget set changes little, and the remaining difference comes from the data.

The original model has been trained on the forget set. This effect is small: on the same held-out nonmembers, the original model and Retrain are at a KS distance of 0.030–0.045 on TOFU and 0.047–0.055 on MUSE-News.

Nonmembers can differ from forget examples in difficulty and domain. We agree that being a nonmember is not sufficient (Sec. 2.1). Even the benchmarks' own held-out nonmembers are not identical to the forget set: under Retrain, the two are at a KS distance of 0.215–0.245 on TOFU and 0.133–0.141 on MUSE-News. When no held-out set exists, we construct the reference with explicit instructions and fixed acceptance rules: TOFU's own attribute-conditioned procedure, and, for MUSE-News, articles that follow the topic, event type, scale, length and paragraph structure of one forget article without repeating its named entities or wording. Prompts and rules: https://anonymous.4open.science/r/ref-construction-release-C364/

"KS to ideal target" is the KS distance between the reference scored by the original model and the forget set scored by Retrain:

BenchmarkReference used by NMLKS to ideal targetLoss gap after NML
TOFU 1BHeld-out0.2580.052
TOFU 1BLLM-constructed0.245–0.3000.077–0.110
MUSE-NewsHeld-out (loss gap from Table 2)0.141–0.1480.123
MUSE-NewsPublic human-written news0.2110.052

On TOFU, the constructed references are as close to the ideal target as the held-out set, and NML with them stays below NPO (0.357), SimNPO (0.237), UnDIAL (0.380) and RMU (0.178); only LUNAR is lower (0.076; MU 0.302). On MUSE-News, public news matches Retrain's loss gap (0.052). Among these references, all within a KS distance of 0.30, the loss gaps do not follow the order of the distances: the distance separates them from a clearly mismatched reference (see (2)) and does not rank them.

(2) Document level versus chunk level. We agree that a match at the document level is not sufficient: a reference should be compared with the forget set at the unit used for training, here the 2048-token chunk. At the document level, the constructed news is as close to the ideal target as held-out news (NLL KS distance 0.156–0.172, against 0.164); at the chunk level it is further away (0.375–0.422, against 0.141–0.148). The cause is that longer constructed articles become more predictable, which the forget articles do not, and chunks give long articles more weight. Even so, NML's loss gap with this reference (0.263–0.331) is below those of NPO and UnDIAL (0.498) and SimNPO (0.455). We will add to the limitations (Sec. 6) that the approximation is reliable when the reference matches the forget set at the training unit.

(3) Assessing a reference without Retrain. We score the forget set and the candidate reference, at the training unit, with a model trained on neither, and compute the KS distance between the two NLL distributions. The check does not need to know why a reference differs: difficulty, domain and scoring unit all act through the NLL distribution, which is what it compares. On MUSE-News, with three scoring models:

MUSE-News referenceKS (checkpoint before fine-tuning)KS (Phi-1.5)KS (Llama-3.2-1B)Loss gap after NML
Public human-written news0.140.050.080.052
Held-out0.20–0.270.09–0.120.11–0.130.123
LLM-constructed0.98–0.991.001.000.263–0.331

LLM-constructed news is at 0.98–1.00 under all three models and has the largest loss gap. On TOFU, the distance is 0.30–0.32 for the held-out reference, 0.43–0.59 when it is fully replaced by QA about real authors, and 0.57–0.65 for QA about fictional companies; the loss gaps after NML are 0.052, 0.255 and 0.216. So, to your question: without Retrain, this check identifies the unsuitable references on both benchmarks: every reference with a final loss gap above 0.2 is clearly further away than the held-out reference of the same task. A milder mismatch is harder to detect but costs less: replacing 20% or 50% of the TOFU reference gives loss gaps of 0.079–0.083 or 0.115–0.120 (reply to Reviewer Beoj, Q2). We have not measured the smallest mismatch the check detects, and present it as a first screening step.

AuthorsOpenReview ↗

Response to Reviewer ueW4 (2/3): W2 & Q3 — relation to earlier work; what energy matching adds

W2 & Q3: relation to earlier work, and what energy matching adds

Relation to earlier work. We agree that calibrating forgetting against unseen data has precedents, and we will add SCRUB+R and KGA to the related work (Sec. 7). They differ from NML in what is calibrated and how:

  • SCRUB+R (Kurmanji et al., 2023; image classifiers): after training, it returns to the checkpoint whose error on the forget set is closest to the error on unseen validation data. It calibrates one number, by choosing a checkpoint.
  • Sula et al. (2025; image classifiers): adds the loss of a separate membership-inference model to the training loss.
  • KGA (Wang et al., 2023): aligns a knowledge gap between models, and needs two auxiliary models.

NML's contribution relative to these methods is that it makes the nonmember NLL distribution the training target itself: it matches the whole distribution of forget NLLs inside the loss, where earlier methods adjust one statistic or rely on auxiliary models, and it does so for LLMs, with a reference that can be held-out or constructed. As a rough comparison, we adapted SCRUB+R to our setting (SCRUB's forget term with our retain term, then its rewinding rule applied to the mean NLL; TOFU 1B, 3 seeds). Since its rewinding step is selected on held-out data, we select on one half of the evaluation examples and report on the other half. There it gives a KS-D of 0.32–0.36 and a loss gap of 0.32–0.34, against 0.081 and 0.073 for NML on the same half (0.063 and 0.052 on the full evaluation, as in the table below).

What energy matching adds beyond a nonmember target (Q3). The nonmember target is what lowers the attack gaps, and matching the whole distribution, as opposed to its mean, is what brings the model closer to Retrain. You are right that mean matching keeps most of the attack-gap improvement; Sec. 5.1 reports the same. For the controlled comparison you ask for, we kept the nonmember reference, the retain term and the schedule fixed, and replaced only the forget term (TOFU 1B, 3 seeds × 2 reference folds):

Forget termKS-D ↓MU ↑Loss gap ↓zlib gap ↓Min-K% gap ↓Min-K%++ gap ↓
Energy (NML)0.0630.6000.0520.0560.0440.114
Mean matching (Table 5)0.1120.5980.0560.0550.0500.106
Wasserstein-10.1000.5990.0470.0680.0400.092
Wasserstein-20.0650.5960.0380.0780.0360.197
RBF-MMD0.1550.5970.1230.0390.1100.135
  • The attack gaps come from the target. Energy, mean matching and both Wasserstein terms all reach loss gaps of 0.038–0.056, against 0.237 for SimNPO (Table 1), which has no nonmember target.
  • Matching the distribution improves forgetting. KS-D, the agreement with Retrain, is 0.063 for energy and 0.112 for mean matching, and energy is lower in each of the 6 runs when the two are compared with the same seed and reference fold. Energy thus improves agreement with Retrain at comparable attack gaps and utility.
  • The improvement does not depend on one particular distance. Wasserstein-2, another distance between distributions, reaches a similar KS-D (0.065). The choice of distance matters: Wasserstein-1 gives KS-D 0.100 and RBF-MMD gives 0.155. We use energy as the default because it has no bandwidth to set, unlike RBF-MMD, and gives lower zlib and Min-K%++ gaps than Wasserstein-2 (0.056 against 0.078, and 0.114 against 0.197) at a slightly higher utility.

We will add the three new forget terms to Table 5, and limit our statement that existing objectives ignore nonmember behavior to the LLM objectives we compare with.

AuthorsOpenReview ↗

Response to Reviewer ueW4 (3/3): W3 & Q1 — MUSE metrics, MUSE-Books, other splits, WMDP

W3 & Q1: MUSE metrics, MUSE-Books, other splits, WMDP

Verbatim memorization and PrivLeak on MUSE-News (3 seeds). We report both metrics as defined in MUSE. A PrivLeak of 0 means that an attack separates forget text from unseen text exactly as well as under Retrain, and values near −100 mean that forget text is still almost perfectly recognizable as training data. Our Full model reproduces MUSE's reported value of −99.8. PrivLeak follows the official MUSE protocol and is computed on MUSE's own forget and holdout excerpts. The audit of Table 2 differs from it in the data and in the scoring unit: its nonmember examples are disjoint from the reference used by NML, and it scores 2048-token chunks. The two are therefore not directly comparable.

MethodVerbMem-F ↓PrivLeak (→ 0)KM-R ↑ (Table 2)
Retrain0.2050.00.551
Full0.571−99.80.546
NPO0.524−99.70.509
SimNPO0.343−80.70.512
UnDIAL0.315−99.80.507
RMU0.295−78.90.501
LUNAR0.093−5.80.277
NML0.232−12.40.537

Under the audit of Table 2, NML has the lowest Loss and zlib gaps; its VerbMem-F, which does not involve the holdout, is the closest to Retrain among the methods that keep retained knowledge. LUNAR's PrivLeak is closer to 0, with roughly half of Retrain's KM-R (0.277 versus 0.551) and less verbatim reproduction than Retrain (VerbMem-F 0.093 versus 0.205). We will add both columns to Table 2.

We also correct a transcription error in LUNAR's row of Table 2: its KM-F is 0.216 (reported 0.416), and its Min-K% and Min-K%++ gaps are 0.055 and 0.124 (reported 0.355 and 0.324). As on TOFU (MU 0.302, Table 1), LUNAR obtains these values at the cost of retained knowledge (KM-R 0.277, against 0.537 for NML). We will correct the row and the corresponding text in the revision.

MUSE-Books (3 seeds, official MUSE evaluation). NML's benefits do not carry over to Books. It removes verbatim memorization (VerbMem-F from 0.997 to 0.108; Retrain 0.144), but its PrivLeak stays at −62.6 (Full: −57.4) and KM-R drops from 0.693 to 0.593; SimNPO obtains a better PrivLeak here (+3.6).

The reason is the structure of this split. The forget set and the holdout are volumes of one book series, so changing the model on the forget volumes also changes its NLL on the holdout volumes. NML takes its target from nonmember NLLs measured before unlearning. It brings the forget excerpts to a median NLL of 4.05, the level of the holdout before unlearning (4.06), but unlearning has meanwhile moved the holdout to 4.36. Forget text therefore remains more likely than holdout text, which is what PrivLeak measures. For comparison, unlearning on News moves the mean NLL of held-out articles only from 1.90 to 1.95. The drop in KM-R has the same origin: the retain set of MUSE-Books is the fan wiki of the same series.

This defines the scope of NML. Its target is valid when unlearning the forget set leaves the model's NLL on the nonmembers nearly unchanged. This holds when the nonmembers share no content with the forget set, such as different authors in TOFU or different articles in News. It does not hold when both are parts of one work: lowering the model's likelihood of some volumes also lowers it on the other volumes. We will state this scope in the limitations (Sec. 6) and add the full Books results, with all baselines, to the appendix.

Other TOFU splits. We repeated the experiment of Table 1 on Forget05 and Forget01 (1B, 3 seeds, same setting). Each cell gives the loss gap / Min-K% gap:

SplitNMLSimNPONPOUnDIAL
Forget050.037 / 0.0270.260 / 0.2780.388 / 0.3730.100 / 0.398
Forget010.054 / 0.0430.308 / 0.3560.423 / 0.4430.215 / 0.262

NML has the lowest loss and Min-K% gaps among the methods shown on both splits, with MU within 0.003 of the listed baselines. This extends the gap reductions in Table 1 beyond Forget10. We will add a discussion of both splits to the appendix.

WMDP. Thank you for the suggestion. NML and its evaluation need two things that WMDP does not provide. First, NML's target and the attack gaps are defined on members and comparable nonmembers, whereas the WMDP forget corpora are not known training examples of the model. Second, no retrained model exists, so agreement with Retrain cannot be measured; WMDP is evaluated by multiple-choice accuracy, which measures removal of a capability and not whether training examples remain identifiable. Our claims are therefore limited to unlearning specified training examples, and we will state this scope in Sec. 6.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision

We thank all reviewers for the discussion. We have uploaded a revised version, with changes marked in blue.

The revision contains the text changes discussed in our responses, including the scope of the method (Sec. 6), the related work (Sec. 7), and a correction to LUNAR's row in Table 2. It also adds several of the additional results from our responses to the appendix (Appendices A–C). The remaining results will be added once they are complete with all baselines.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.