acceptodds
Under review as a conference paper at ICLR 2027

Learning What to Forget: Distributional Unlearning for LLM Representation Spaces

Abstract

Machine learning systems increasingly face the need to remove the influence of entire data domains, such as toxic language, harmful behavior, or topical content, rather than isolated records. Recent work formalizes this problem as distributional unlearning: selecting a subset of a forget domain whose removal moves the training distribution away from an unwanted population while preserving proximity to the desired one. However, existing analyses often impose parametric assumptions to obtain tractable selection rules. These assumptions may be poorly suited to high-dimensional language-model representations. We introduce Mamushi, a framework for non-parametric distributional unlearning that ranks forget examples using a probabilistic classifier whose Bayes-optimal logit equals the forget-to-retain log-density ratio (up to an additive class-prior constant). We show that thresholding the population log-density ratio yields the optimal fixed-budget selection rule for our removal–preservation objective and establish a non-asymptotic transfer guarantee relating score-estimation and threshold-calibration errors to degradation from the population-optimal selection rule. Our empirical evaluation spans real-world datasets on toxic-language removal and topical-domain removal regimes using different representations, with Mamushi achieving a more favorable removal–preservation trade-off than other baselines. Our work shows that Mamushi can serve as an efficient selection approach for downstream machine unlearning procedures, reducing the number of forget examples required to reach a fixed forgetting target.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Distributional Unlearning, Data Selection, Density-Ratio Estimation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_LG5HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the problem of distributional machine unlearning where the goal is to sufficiently unlearn the forget distribution while remaining close to the retain distribution (instead of unlearning individual data records). Existing distributional machine unlearning work may be poorly suited for the LLM embedding spaces as they enforce restrictive parametric models. The paper proposes a new framework MAMUSHI that prescribes how to select a forget subset to maximise KL divergence from the forget set distribution and minimise divergence from the retain set distribution. MAMUSHI is non-parametric and selects forget data with the largest logits based on a classifier separating the forget and retain data. The paper presents theoretical results that bound the KL divergence based on the log-density and threshold calibration errors. The empirical results on various datasets show that MAMUSHI minimise the absolute difference in perplexity on smaller deletion budgets.

Critical strengths
  1. The paper is mostly well-written and easy to follow. The solution, theoretical assumptions and results, and differences with existing work are clearly explained.
Critical weaknesses
  1. The key contribution of the paper is to improve the forget-utility tradeoff under a limited deletion budget. However, the scenario/significance of a limited decision budget is not sufficiently motivated or explained.
  2. The experiments in the main paper are limited. Appendix D includes some additional experiments and it would be helpful to summarise them in the main paper.
Questions for authors
  1. In practice, why would a smaller deletion budget occur and why is data selection for unlearning important? Data selection or pruning during learning can result in better performance than using the whole dataset. Can data selection for unlearning result in better performance than using the whole forget set too?
  2. Can you add some ablation studies on how Mamushi performance depends on the accuracy/BCE loss of the classifier? The classifier might be less accurate when (a) a less suitable model is used/insufficient training or (b) the forget and retain distribution may be more similar.
  3. Some works, such as [C], have discussed that density ratios may be difficult or inaccurate to estimate for high-dimensional data, especially when the densities of interest are sufficiently different (e.g., when the forget and retain distribution are very separated). Discuss the impact on MAMUSHI.

[C] Choi, K., Liao, M., & Ermon, S. (2021, December). Featurized density ratio estimation. In Uncertainty in Artificial Intelligence (pp. 172-182). PMLR.

5 comments
AuthorsOpenReview ↗

Reviewer LG5H ( Rebuttal I)

We thank the reviewer for their careful and constructive feedback. We address each point individually below, and have updated the manuscript accordingly (major changes marked in blue)

W1/Q1 Motivation for a limited deletion budget

We agree this should be motivated more explicitly in the introduction and have added a paragraph to that effect. In short, budgeted selection matters for three reasons:

(i) Compute. The cost of efficient unlearning methods scales with the forget-set size, since gradient-based, saliency-based, influence-function, and noise-based certified methods all compute per forget sample (Guo et al., 2020; Kurmanji et al., 2023; Fan et al., 2024; Allouah et al., 2026, Remark 11). The budget is therefore typically set by the compute available downstream.

(ii) Utility. When the forget domain is intertwined with the retained one, full deletion can harm utility, as shown by Allouah et al. (2026, Sec. 4.2.2).

(iii) Annotation. A small effective subset reduces the need to exhaustively identify the forget domain.

Can selecting a subset outperform deleting the entire forget set (as data pruning can beat full-data training)? Yes, but here the axis matters. On distributional matching our SAD metric measures distance to the gold model (retrained on the retain set alone), so deleting the whole forget set is optimal by construction (SAD→0\mathrm{SAD} \to 0 at β=1\beta{=}1); selection’s value there is reaching near-oracle removal with a small fraction of the deletions (Fig.2-3) — the direct analogue of pruning matching full-data performance at lower cost. On utility, selection can strictly dominate: the full-deletion (gold) model discards the entire forget domain, including benign in-domain content, and when the domains are intertwined this over-corrects and harms utility (Allouah et al. 2026, Sec. 4.2.2); a targeted subset removes the distributionally potent samples while preserving the benign remainder, yielding a better removal–utility trade-off than deleting everything. This is precisely what our two-sided objective formalizes — Lemma 5.1 splits the edit into removal and preservation terms, and the threshold selector (Prop. 5.2) optimizes their trade-off at each budget rather than maximizing removal alone. In short, selective forgetting is the unlearning counterpart of “less-is-more” data pruning: it attains the oracle’s removal efficiently, and under intertwined domains it can exceed full deletion on utility.

W2. Reference Ablations and other theoretical results from Main Text

We agree and have added a summary of Appendix B and D to Section 5 and 6.

Q2 Classifier Ablation

Ablation on Classifier choice (model class) is already included in the manuscript. MLP and LogReg have essentially identical out-of-fold AUROC (0.971 vs. 0.965) and similar log loss (0.191 vs. 0.210), yet the MLP yields lower SAD across budgets. This shows that standard classification metrics are not reliable proxies for selection quality, for downstream effect. They measure how well forget samples are separated from retain samples on average (pure classification), whereas MAMUSHI cares about how the estimated log-density ratio orders forget samples among themselves. Estimator flexibility matters precisely for this finer-grained ordering, which motivates our use of an MLP.

a) Regarding training time, we noticed our MLP reaches its validation optimum with early stopping approximately on average at 12 iterations. We therefore capped training max-limit to half this horizon ( 66 iterations), which captures the validation-optimal model while halving the compute. The undertrained MLP MLP selector attains higher SAD than original setting, raising mean SAD from 3.463.46 to 4.914.91 (a ≈42%\approx 42\% increase). This indicates that training the density-ratio estimator to its early-stopping horizon is helps with selection quality.

b) We first note that in real deployments the separation between the forget (P1P_1) and retain (P2P_2) distributions is intrinsic, not a parameter we control: it is fixed by the embedding geometry and by the semantic content that defines the forget domain. By construction that domain carries characteristic structure — toxic slurs and profanity for Jigsaw (Table 4), or guns/war/destruction/politics vocabulary for the topical setting — so a non-trivial P1P_1–P2P_2 difference is always present and cannot be artificially reduced on real data. To nonetheless characterize behavior as the two distributions approach each other, we run a controlled synthetic study: heavy-tailed (tt, df$=4$), anisotropic features in R50\mathbb{R}^{50}, with P1P_1 shifted from P2P_2 by a controllable amount (sep$=0$ means identical), scored by the same cross-fitted MLP used in the paper.

sepOOF AUCDFD_Fspreadfrac ∣ℓ∣<1\lvert\ell\rvert<1
0.00.5030.000.380.97
1.00.8211.211.720.34
2.00.9552.472.440.10
3.00.9863.622.870.04
AuthorsOpenReview ↗

Reviewer LG5H ( Rebuttal II)

i) Graceful degradation with a safety floor: as P1→P2P_1 \to P_2 the learned logits compress toward zero (spread 2.87→0.382.87 \to 0.38, DF→0D_F \to 0), and at exact identity the scorer collapses to chance (AUC 0.500.50) with 97%97\% of logits in ∣ℓ∣<1|\ell|<1 — i.e. it reports no structure and induces no spurious selection, which is the desired behavior.

(ii) The margin density grows with overlap: the fraction of mass near the threshold rises monotonically as the distributions approach, which is exactly the large-MM regime our Corollary~5.7 identifies as the hard case where accurate ℓ^\widehat\ell matters most.

(iii) Our real experiments operate in the healthy intermediate band: Jigsaw yields DF≈2.2D_F\approx2.2, AUC$\approx0.96$, matching sep$\approx2$ above — meaningful separation, neither degenerate nor saturated. The method thus behaves sensibly across the full overlap spectrum and degrades gracefully rather than catastrophically.

Q3 New Reference

We thank the reviewer for pointing to [C]. Their analysis shows that, when the two densities are far apart, a probabilistic classifier trained on finite samples tends to inflate its logits (e.g., w1→∞w_1 \to \infty in their Sec. 3.1), producing inaccurate ratio values. This matters for the tasks they target (importance weighting, mutual information, importance resampling), which rely on ratio magnitudes. MAMUSHI is less exposed for three reasons:

(i) Selection uses only ranking near the threshold. MAMUSHI relies on the ranking of forget samples and the deletion threshold, not on ratio values. The pathology in [C] largely rescales the logit, which preserves the ordering. More generally, Theorem 5.6 shows that only score errors near the threshold affect selection.

(ii) Separated domains are where selection is most robust. Corollary B.3 shows that when little score mass lies near the threshold, as for well-separated domains, the selection gap due to estimation error vanishes.

(iii) Our settings are not in the separated regime. The forget and retain scores overlap substantially (Fisher discriminability ≈ 2.5, calibrated scores; App. C.2), and the classifier is regularized and early-stopped.

Unlike [C], which requires an invertible map so that feature-space ratios equal input-space ratios, our objective and guarantees are defined directly in the representation space, so ratio preservation is not needed. Since our guarantees are estimator-agnostic, featurized [C] or telescoping (Rhodes et al., 2020) estimators can be used directly under extreme separation. We have added this discussion to the paper.


We thank the reviewer for carefully reading our work and hope our response addresses the concerns raised, and we are happy to clarify any remaining points. If they feel their concerns have been resolved, we would be grateful if they would consider updating their assessment.

Thank you for the detailed response! My concerns have been mostly resolved.

In the response, the authors commented

We first note that in real deployments the separation between the forget and retain distributions is intrinsic, not a parameter we control: it is fixed by the embedding geometry and by the semantic content that defines the forget domain.

I understand the point. My initial question can be rephrased as: Does MAMUSHI work for other real deployment scenario where the separation between forget and retain distribution might be much lower or higher (e.g., by data owner instead of topic)?

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

This paper proposes MAMUSHI, a data selection method for distributional unlearning. It trains a classifier on LLM embeddings to estimate the density ratio between forget and retain data. Samples with higher scores are selected for removal. The authors provide an optimality result and an error bound. Experiments on two datasets and three LLMs show better removal results than several baselines.

Critical strengths
  1. Useful problem setting. Selecting a small set of important samples for unlearning is practical. It may reduce the cost of removing unwanted data.

  2. Good empirical results. MAMUSHI performs well across three LLMs and two datasets. The authors also test it with existing unlearning methods, such as NegGrad+ and SalUn.

Critical weaknesses
  1. The paper assumes that samples with higher density ratios are more important to forget. However, being typical of the forget distribution does not necessarily mean having more influence on the trained model. The paper does not clearly establish this link.

  2. Allouah et al. (2026) already studied selective data removal for distributional unlearning. Menon and Ong (2016) showed how classifiers can estimate density ratios. This paper mainly combines these existing ideas with LLM embeddings. Comparisons with other non-parametric density-ratio estimators are also missing.

  3. The optimality result only applies to the edited data distribution, not the actual unlearning behavior of the model. Also, Theorem 5.6 assumes bounds on score and threshold errors but does not show how these errors depend on the training sample size. The practical meaning of the guarantee is therefore unclear.

  4. Eq. (7) calls SAD a sum of absolute distances, but the formula contains no absolute values. Changes in forget and retain perplexity may cancel each other. A lower SAD may therefore not always mean better unlearning.

Questions for authors
  1. Why should a high density ratio indicate greater influence on the trained model? Can the authors provide direct evidence for this assumption?

  2. What is the main novelty beyond Allouah et al. (2026) and Menon and Ong (2016)? Can the authors compare against other non-parametric density-ratio estimators?

  3. Can the theoretical guarantees be extended to actual model unlearning? Can the authors provide a sample-size-dependent bound for the learned classifier?

  4. Does Eq. (7) miss absolute value signs? If not, how do the authors prevent the two perplexity terms from canceling each other?

3 comments
AuthorsOpenReview ↗

Reviewer GbiZ ( Rebuttal I)

We thank the reviewer for their careful and constructive feedback. We address each point individually below, and have updated the manuscript accordingly (major changes marked in blue).

W1/Q1 High Density Ratio & Trained Model

We thank the reviewer — this is an important point, and it lets us correct a subtle but central distinction. MAMUSHI does not rank by typicality under the forget distribution; it ranks by the density ratio ℓ∗(x)=log⁡p1(x)p2(x)\ell^*(x)=\log \frac{p_1(x)}{p_2(x)}. A point that is typical of the forget set but equally typical of the retain set has a low ratio and is not selected. So “typical of p1p_1” and “high ℓ∗\ell^*” are different quantities, and the reviewer’s distinction is exactly the one our score is built to respect: we target points that are distinctively forget-versus-retain, not merely frequent under p1p_1. This is also consistent with the data-centric framing we share with Allouah et al. (2026): the selection rule is estimator-agnostic by design, and Prop. B.12 now makes its effect on the retrained model explicit.

On the explicit link to the trained model, we agree a direct result adds completeness. Following the spirit of Prop. 2 in Allouah et al. (2026), we add Proposition B.12 (Appendix, referenced from the main text). For a model retrained on the edited data, the forget/retain log-loss gaps to the gold-standard model are exactly.

Δ1(S)=a(β)−∫Scp1 ϕ(ℓ∗),Δ2(S)=a(β)−∫Scp2 ϕ(ℓ∗),a(β)=log⁡(1+π1π2(1−β)).\Delta_1(S)=a(\beta)-\int_{S^c} p_1\,\phi(\ell^*),\quad \Delta_2(S)=a(\beta)-\int_{S^c} p_2\,\phi(\ell^*),\quad a(\beta)=\log\big(1+\tfrac{\pi_1}{\pi_2}(1-\beta)\big).

Since ϕ\phi is increasing in ℓ∗\ell^* and a(β)↓0a(\beta)\downarrow0 at β=1\beta{=}1, removing high-ratio points reduces these gaps the most — i.e. influence on the trained model is monotone in the density ratio, which is precisely the assumption the reviewer asked us to justify.

Direct evidence:

  1. Ratio ordering is causal, not incidental: removing the lowest-ratio points first (anti-MAMUSHI) does worse than random removal (App. D.2, Fig. 5).
  2. Typicality is a strictly worse proxy than the ratio: CORESET, which ranks by typicality under p1p_1 irrespective of the ratio, needs a 64% budget to reach the target MAMUSHI reaches at 24% under retraining (Table 1).
  3. The thresholded set is optimal at every budget: MAMUSHI attains the lowest SAD across budgets (Figs. 2–3), with the gap shrinking to 00 as β→1\beta\to1 (non-strictly), consistent with the envelope in Prop. B.12.

Q2. Novelty

We thank the reviewer for giving us an opportunity to explain the novelty in our work. We would like to highlight the following points.

  1. Linking DU and DRE: Allouah et al. introduce and frame distributional unlearning and analyze it primarily under Gaussian assumptions. Their section 5 identifies extending the analysis beyond Gaussians, and applying it to LLM representations, as open directions. Their practical rules are centroid contrasts (LR-COS, LR-MAHA), which assume a geometry that LLM embeddings do not have: both domains are strongly anisotropic with distinct principal subspaces (Appendix D.5). We tackle this problem in our paper, where simple selection rules are unsuitable given the geometry. Our contribution is to show that, selective distributional unlearning( without any parametric assumption) is indeed a density-ratio problem (DRE). We practically learn the ratio using seminal link between DRE and Binary Classification ( Menon & Ong.)

From our manuscript:

"Specifically, we extend selective distributional unlearning to high-dimensional LLM embeddings, replacing parametric distance scores with an estimated log-density ratio that makes no assumption on the distributions through the lens of statistical decision theory, reducing density-ratio estimation to standard binary classification."

We hope the reliance on existing ideas is not seen as a forced combination, but rather identifying a failure mode and correcting it as required.

  1. On the theoretical side to our knowledge the following results are new:

(i) Optimal selection. Allouah et al. characterize the removal–preservation frontier over distributions in a Gaussian or exponential family. Their selection rules are heuristics, and they observe that none dominates the others (their App. B.2.3). We show that, with no distributional assumption, thresholding ℓ∗\ell^* is the optimal deletion set, simultaneously for removal and preservation, at every budget (Lemma 5.1, Prop. 5.2).

(ii) Estimator-agnostic transfer. Their guarantees cover one specific rule under univariate Gaussians. Theorem 5.6 instead bounds the loss incurred by any estimated score, through its score and threshold errors.

AuthorsOpenReview ↗

Reviewer GbiZ ( Rebuttal II)

(iii) Guarantee through the training loss. Menon & Ong show that BCE regret controls density-ratio estimation error. We carry this through to the unlearning objective: Corollary B.1 bounds the gap to the optimal selector by the classifier’s BCE excess risk, so the quantity minimized in practice directly controls unlearning quality.

(iv) Model-level link (New). Their Prop. 2 bounds the log-loss of a conditional predictor for a given edited distribution. Our new Prop. B.12 gives exact log-loss gaps for a generative model retrained on the edited data, at every budget, and shows that the thresholded set is also optimal at the model level.

Other than motivational and algorithmic novelty, on the experimental end extending distributional unlearning to generative LLMs for text data requires change in the original setup (synthetic Gaussians and TF-IDF features). Since the end goal is to remove the influence of the forget data on a generative model, we measure unlearning through held-out forget and retain perplexity, the model’s native log-loss, rather than through accuracy or F1 from an external classifier. We measure it relative to a gold-standard model fine-tuned on retain data alone, which links the evaluation directly to our KL objectives (now Prop. B.9). Selection operates in a representation learned from a disjoint half of the data and shared by all methods, so it is leak-free and identical across them. Every point on our curves requires fine-tuning a fresh LLM, repeated across 3 model families, 2 removal regimes, and 5 seeds, 9 budgets, and across 8 selection methods.

Q2. Comparison with different density-ratio estimator

We are currently running experiments for two additional non-parametric density-ratio baselines (kk-NN and RuLSIF) across all budgets, datasets, embeddings, and seeds. Because each data point requires fine-tuning a fresh LLM, these runs take substantial compute time, and we will update this response with the full results as soon as they finish before the rebuttal deadline.

Q3. Actual Model Unlearning and Finite Sample Unlearning

Model level-unlearning. Our paper's core contribution lies in extending Distributional Unlearning to LLM embeddings where simplified geometric assumptions fail.

From the manuscript,

"What is the minimal set of data points to remove from the forget set for maximal distributional impact where data is natural language?"

Like Allouah et al. (2026), we stick to the framing of holding the learning or unlearning procedure fixed and asks how the choice of data to remove changes the outcome (Motivational Figure 1).

The interaction with specific unlearning algorithms is evaluated empirically (Table 1) as a secondary-parallel branch. Guarantees for specific approximate unlearning algorithms (e.g., NegGrad+, SalUn) are inherently algorithm-dependent, and they form a natural future research direction of their own.

Sample-size dependence. We had noted this as future work (Sec. 7) because the score error, and hence any sample-size dependence, is determined by the probabilistic classifier chosen, while MAMUSHI is a selection principle that does not depend on that choice. Remark B.2 already notes that any generalization bound ∥ℓ^−ℓ∗∥L1(p1)≤rn(ρ)\|\hat\ell-\ell^*\|_{L^1(p_1)}\le r_n(\rho) directly yields an (α,ε)(\alpha,\varepsilon) guarantee through Corollary B.1, and Corollary B.8 provides one such route via the BCE excess risk. Since the reviewer asks, we make it explicit.

In Appendix we reduce the score error to the classifier’s BCE excess risk (Corollary B.8), bounding the gap to the optimal selector by O(regBCE14)O(\mathrm{reg}_{\mathrm{BCE}}^{\frac{1}{4}}).

Any excess-risk rate for the chosen model class thus yields a sample-size-dependent bound. For example, for ReLU networks trained with logistic loss, Zhang, Shi & Zhou (JMLR 2024) show regBCE=O~(n−s/(s+d))\mathrm{reg}_{\mathrm{BCE}} = \tilde O(n^{-s/(s+d)}) when the class probability Pr⁡(Y=1∣x)\Pr(Y=1\mid x) is Hölder-ss smooth, where dd is the input (embedding) dimension. This implies a gap of O~(n−s/(4(s+d)))\tilde O(n^{-s/(4(s+d))}), with n=n1+n2n = n_1+n_2 the classifier’s training size. We do not claim any new statistical learning theoretic results and have added this as a remark after Corollary B.8.

Q4. SAD Typesetting

Eq. (7) does include absolute values, so the two terms cannot cancel. The large delimiters may not have rendered clearly in some PDF viewers; we have revised the typesetting to make them unambiguous.


We thank the reviewer for carefully reading our work and hope our response addresses the concerns raised, and we are happy to clarify any remaining points. If they feel their concerns have been resolved, we would be grateful if they would consider updating their assessment.

Dear Reviewer GbiZ,

We thank you for your time and detailed review. In response, we added a proposition linking the selection to the retrained model (Prop. B.9), a sample-size-dependent guarantee for the learned classifier (Remark B.8), non-parametric density-ratio baselines (KNN-Ratio, RuLSIF), and clarified the absolute values in Eq. (7). As the discussion period closes on November 18, we would be grateful if you would consider updating your score in light of the revisions. If any concern remains, we would be happy to provide further clarification.

Best,

The Authors

Reviewer_st2pOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes a method to unlearn a probability distribution (forgotten) from an original data distribution. The target is an unlearned data distribution that has minimised KL divergence with the retained distribution and maximised KL divergence with the forgotten distribution. In other words, the contamination of the original data by the forgotten data is removed.

The used technique is a calibrated thresholding of the density log ratio similar to the likelihood ratio test (LRT) tricks in order to identify a set of forgotten data samples SS whose probability under the forgotten density equals the available unlearning budget β\beta. The authors show that their technique identifies an optimal set in the sense of their objective, i.e., KL minimisation-maximisation among all measurable sets with the same budget.

The authors validate the efficacy of their proposed technique by showing that MAMUSHI (their method) requires less budget β\beta (less number of forgotten samples) to achieve a specific sum of absolute distances in terms of perplexity between the fully retrained model and the model retrained without the chosen forgotten samples, demonstrating that MAMUSHI gives the lowest budget among the presented baselines for the same level of sum of absolute distances.

Critical strengths

S1. The paper covers a clear gap in prior work as it extends distributional unlearning to the case of general data distributions. The exposition is clear and easy to follow. All notations are defined precisely and used consistently. The motivation, contribution, and treatment are highlighted sufficiently.

Critical weaknesses

Although the paper is easy to follow with a clear objective, my major concern is related to the significance of the presented material which make me recommend rejection of the paper in its current form.

W1 [CRITICAL]. The paper fails to highlight the significance of the proposed approach. The whole idea seems to be of a binary classifier trained to classify a given data point as a retained or forgotten data and transform the data distribution into a distribution that diverges from the forgotten distribution and aligns with the retained distribution. This is a special case of unsupervised domain adaptation (where the new domain is a restricted subdomain of the original one; i.e., the retained domain) and the exact problem of demixing/decontamination. General well known techniques to which the proposed framework holds high resemblance include density ratio estimation, mixture proportion estimation, and the learning from positive and unlabelled data problem (which arguably requires less information). Hence, introducing learned density-ratio selection as one of the contributions is not justified. Lemma 4.1 requires a citation since it is a well known result of binary classification. Thresholding the log probability ratio is a common technique in binary hypothesis testing known as the likelihood ratio test (LRT), and calibrating the threshold so as to achieve a particular level β\beta is just the same as calibrating the threshold in LRT to achieve a particular type I error of level β\beta. That trick appears in Neyman and Pearson (1933). At the present time, all these results are considered trivial, which makes the contribution of the paper more of an engineering application of established techniques. Moreover, there is nothing in the paper that links this idea to practical machine unlearning and tells us what is expected to happen for a model (possibly a deep learning model) trained on the original data distribution when we try to unlearn the forgotten distribution (subpopulation), or how to do that. I feel that the paper needs major improvement to highlight the significance of the proposed technique, given the rich literature on this problem.

Questions for authors

Q1. Why is the i.i.d. assumption separately within p1p_1 and p2p_2 introduced in line 159 needed?

Q2. Assumption 5.3 is unpractical. What would happen in the regions of high separability between p1p_1 and p2p_2, where the log ratio goes to ∞\infty?

Q3. The error in calibration rτr_{\tau} is directly linked to the error in the score estimation ∥∥ℓ^−ℓ∗∥∥\|\|\hat{\ell} - \ell^{*}\|\| since rτr_{\tau} is directly related to the cumulative distribution function of the random variable (ℓ^−ℓ∗)(x)(\hat{\ell} - \ell^{*})(x). Is it possible to simplify the errors in Theorem 5.6 by exploiting such a relation? This would make your statement in line 347 (hence the guarantee) concern only one error instead of two.

Q4. Table 1 shows that the gain in deletion budget highly depends on the unlearning method: for instance, Coreset requires less budget than MAMUSHI when SalUn is used. The difference between Coreset and MAMUSHI seems to be within statistical error when NegGrad+ is used. Please explain this dependence on the unlearning method and include standard deviation in Table 1.

7 comments
AuthorsOpenReview ↗

Reviewer st2p ( Rebuttal I)

We thank the reviewer for their careful and constructive feedback and for recognizing that the paper addresses a clear gap and is clearly presented. We value you highlighting connections to classical statistics and related learning problems, which has helped us position the contribution more precisely. We address each point individually below, and have updated the manuscript accordingly (major changes marked in blue).

W1. Significance and relation to prior work

  1. “A binary classifier trained to classify a point as retained or forgotten … a special case of unsupervised domain adaptation.”

We agree that MAMUSHI uses a forget-versus-retain classifier. However, we respectfully clarify, that the problem is not a special case of unsupervised domain adaptation:

Inputs: Domain adaptation has labeled source data and an unlabeled target. In distributional unlearning, both forget and retain samples are labeled.Output: Domain adaptation produces a predictor that performs well on the target. Our output is an edit to the training set. The budget: If the problem were treated as adapting to the retain domain, the solution would be to discard or downweight all forget data, i.e., β = 1. The deletion budget is what makes the problem non-trivial: deciding which forget samples to remove under a fixed budget, trading off removal against preservation. Allouah et al. (2026) draw the same distinction. We have added this discussion to the related work (Sec. 2) in Appendix.

  1. “The exact problem of demixing/decontamination.”

We respectfully clarify that the problems differ. Decontamination and mixture proportion estimation recover the components or mixing proportions of contaminated distributions from unlabeled mixtures (Blanchard et al., 2010; Scott et al., 2013; Ramaswamy et al., 2016; Katz-Samuels et al., 2019), and identifiability is central there. In our setting, the forget samples are already identified. We do not recover any distribution; we select which identified samples to delete under a budget, and we evaluate the effect on a retrained model. The shared ingredient is the density ratio, which we now discuss explicitly with these references in Sec. 2.

  1. “Resemblance to density ratio estimation, mixture proportion estimation, and PU learning (which arguably requires less information).”

We agree on the shared density-ratio machinery. We have added a related-work paragraph that positions MAMUSHI relative to domain adaptation, mixture proportion estimation and decontamination, and PU learning (du Plessis et al., 2014; Kiryo et al., 2017; Bekker & Davis, 2020). Our assumption of labeled forget and retain samples follows the established formulation of distributional unlearning, where the forget domain is identified by an upstream process (Allouah et al., 2026, Sec. 2). Relaxing it to a labeled forget sample and an unlabeled training pool is a future direction and we now discuss it in the conclusion.

We are happy to include any specific references, other than the new ones we already included, the reviewer feels would strengthen the positioning of the manuscript.

  1. “Introducing learned density-ratio selection as a contribution is not justified.” “Lemma 4.1 requires a citation.”

The classifier-based density-ratio estimator is due to prior work (Menon & Ong, 2016; Sugiyama et al., 2012), and we did not intend to claim it. We have reworded the contribution as follows: budgeted distributional unlearning reduces to thresholding the log-density ratio, which can be estimated with any probabilistic classifier, without parametric assumptions. Lemma 4.1 is now attributed to Menon & Ong (2016) and Sugiyama et al. (2012). We note that it is stated for completeness.

  1. “Thresholding the log-ratio is the likelihood ratio test; calibrating the threshold is the Neyman–Pearson trick.”

We now state this explicitly: the oracle selector of Prop. 5.2 is a Neyman–Pearson-type region (Neyman & Pearson, 1933). Sec. 3 already presented the method within the likelihood-ratio principle (Neyman & Pearson, 1933; Wald, 1992), and the revision makes the connection explicit at Prop. 5.2.

AuthorsOpenReview ↗

Reviewer st2p ( Rebuttal II)

Having made the connection explicit, we respectfully note what the analysis adds. The Neyman–Pearson lemma optimizes a functional that is linear in the set, subject to a linear constraint. Our objectives are KL divergences between the forget or retain distribution and a renormalized edited mixture, so they are nonlinear in the selected set. Lemma 5.1 shows that both nonlinear objectives reduce to integrals of the same increasing function of ℓ∗\ell^*. That is what makes a Neyman–Pearson-type argument applicable, and it yields a single set that is simultaneously optimal for removal and preservation. To our knowledge, this characterization is new to distributional unlearning. The closest prior work analyzes the problem only under Gaussian assumptions, with heuristic selection rules, and observes that no rule dominates the others (Allouah et al., 2026, App. B.2.3).

  1. “All these results are considered trivial, making the contribution an engineering application of established techniques.”. “Nothing in the paper links this idea to practical machine unlearning ”

We appreciate this perspective and agree that our results are built on classical tools from statistics and probability. We respectfully suggest that the value of a result proved with standard techniques lies in what it shows (how it connects different fields) rather than in the difficulty of its proof. Here, the analysis shows that budgeted distributional unlearning, despite being posed through two nonlinear KL objectives on a renormalized mixture, reduces exactly to a density-ratio estimation problem: at any fixed budget, the removal–preservation trade-off studied in prior work collapses to a single optimal set determined by the log-density ratio. This connects distributional unlearning to the classsifier-based density-ratio literature the reviewer highlighted previosuly. This is correctly placed in the problem as previous geometric selection are not suitable for our setting. We see the simplicity of understanding our proofs as a key strength. We believe this connection, together with the first evaluation of distributional unlearning on generative LLMs, will be of interest to the community.

On practicality, we respectfully note that the experiments already evaluate models directly: Sec. 6.1 measures retrained LLMs against a gold-standard model( the bottom-right the red star in Figs 2-3), Sec. 6.2 pairs selection with popular unlearning algorithms NegGrad+ and SalUn, and App. D.5 shows generations.


We believe the revised positioning, the corrected attributions, and the new model-level and finite-sample results address the concern about significance, and we would be grateful if the reviewer would reconsider the paper in light of them.

2. Q1

We follow the problem statement on page 3 of Allouah et al. (2026), in which the forget and retain samples are i.i.d. draws from their respective populations p1p_1 and p2p_2. The assumption is stated separately because the two sets come from different distributions. Samples are identically distributed within each set, not across them. This is the standard two-sample setting of density-ratio estimation (Sugiyama et al., 2012).

It is needed because MAMUSHI estimates ℓ∗=log⁡(p1/p2)\ell^* = \log(p_1/p_2) from these samples. Lemma 4.1 identifies ℓ∗\ell^* with the Bayes-optimal logit for labeled data generated as Y∼Bernoulli(π1)Y \sim \mathrm{Bernoulli}(\pi_1), X∣Y∼pYX \mid Y \sim p_Y. The pooled, labeled set S1∪S2S_1 \cup S_2 is a sample from this model only if each SiS_i consists of independent draws from pip_i. The same condition underlies the generalization arguments through which the score error in Theorem 5.6 is controlled (Remark B.2, Corollary B.8), and it makes our cross-fitted, out-of-fold scores honest estimates of ℓ∗\ell^*. Our new finite-sample guarantees (Remark B.10) follow suit.

3. Q2

We clarify the role of Assumption 5.3 below.

Where it enters the theory . Boundedness is used only to bound ϕ(ℓ∗)≤ϕB\phi(\ell^*) \le \phi_B in the removal guarantee (Theorem 5.6, Step 3). The preservation guarantee does not depend on BB: substituting Δrem\Delta_{\mathrm{rem}} into Eq. (6) gives Δpres(δ)=π1π2[a/δ+2M(δ+rτ)]\Delta_{\mathrm{pres}}(\delta) = \tfrac{\pi_1}{\pi_2}\big[a/\delta + 2M(\delta + r_\tau)\big], which uses only the mutual absolute continuity in Assumption 5.3 (Step 4). Assumptions of this kind are common in analyses of importance weighting and density-ratio estimation (Cortes et al., 2010; Sugiyama et al., 2012).

Regions of high separability . Points where ℓ∗\ell^* is very large lie deep in the forget region, far above the threshold, so both the oracle and the estimated selector remove them. They affect the guarantee only if score error places them on the wrong side of the threshold. In addition, Corollary B.3 shows that when little score mass lies near the threshold, as happens for well-separated domains, the selection gap vanishes. Separability therefore makes selection more robust, not less.

AuthorsOpenReview ↗

Reviewer st2p ( Rebuttal III)

In practice learned logits are bounded. The classifier is ℓ2\ell_2-regularized and early-stopped, and the cross-fitted scores lie in a finite range (e.g., [−10.9,22.7][-10.9, 22.7] on Jigsaw with Gemma embeddings; App. C.2). Moreover, MAMUSHI uses only the ordering of the scores, so clipping or any monotone rescaling of the logits leaves the selected set unchanged. Assumption 5.3 is therefore not restrictive for the selection rule as used.

4. Q3

We thank the reviewer for this insightful suggestion, which is a valuable addition to our manuscript. When the threshold is calibrated at the population level, τ^\hat\tau is the (1−β)(1-\beta)-quantile of ℓ^(X)\hat\ell(X). Its deviation from τ∗\tau^* is therefore governed by the score error through the distribution function of the score. We make this precise in a new corollary in our manuscript. Under a two-sided version of the margin condition, in which the density of ℓ∗(X)\ell^*(X) near τ∗\tau^* is also bounded below by m>0m>0, we show that ∣τ^−τ∗∣≤δ+a/(mδ)|\hat\tau-\tau^*|\le\delta+a/(m\delta) with a=∥ℓ^−ℓ∗∥L1(p1)a=\|\hat\ell-\ell^*\|_{L^1(p_1)}. Substituting this into Theorem 5.6 yields a guarantee in terms of a single error,

α^≥α∗−4 ϕBM(1+2M/m) a,ε^≤ε∗+4M(1+2M/m) a.\hat\alpha\ge\alpha^*-4\,\phi_B\sqrt{M(1+2M/m)\,a},\qquad \hat\varepsilon\le\varepsilon^*+4\sqrt{M(1+2M/m)\,a}.

We retain the two-term form in Theorem 5.6 because it holds without the additional lower bound on the score density, being more flexible. We have kept the line as it is, but added this extension as a corollary in teh appendix, merging both the error sources.

5. Q4

We agree that standard deviations are needed to interpret the close entries in Table 1, and we are currently populating these results over seeds. This is still in progress; it may take some time, but we will report mean ± standard error for every entry and update the manuscript accordingly before the rebuttal deadline. We appreciate the reviewer’s patience. In the meantime, we explain the dependence on the unlearning method, which is expected and reflects how much each method’s outcome depends on which samples are selected.

Retraining is purely data-centric: the outcome is determined entirely by the samples that remain, which is the setting our analysis addresses (Prop. 5.2, Prop. B.9). Here MAMUSHI’s advantage is largest (24% vs. 62–64%), and CORESET — which removes the samples most typical of p1p_1 regardless of the retain distribution — performs no better than random. This is a large margin that we expect to be robust to seed variation.

NegGrad+ performs gradient ascent on the selected samples. Prototypical forget samples, which CORESET tends to select, induce large, coherent ascent directions toward the forget concept; we hypothesize this is why CORESET improves sharply under NegGrad+ (64% → 26%) and approaches MAMUSHI, which remains lowest (21%). The point estimates here are close, and we will report whether the remaining gap is within statistical error once the seed runs complete.

SalUn masks the parameters most salient to the forget loss and retrains them with random labels, which removes the forget concept from almost any subset of p1p_1. Selection therefore matters least under SalUn: all rules perform similarly and even random selection needs only 40%. The MAMUSHI and CORESET point estimates differ by one percentage point (36% vs. 35%), which we expect to fall within statistical error; we will confirm this with the seeded results.

This pattern — savings from selection varying with the unlearning method and being smallest for methods that already approximate retraining well — matches Allouah et al. (2026, Table 5). Across methods the point estimates show MAMUSHI as best or essentially tied, with its advantage largest where unlearning depends most on the data; we will substantiate the ties with standard errors in the updated Table 1.


We thank the reviewer for carefully reading our work and hope our response addresses the concerns raised, and we are happy to clarify any remaining points. If they feel their concerns have been resolved, we would be grateful if they would consider updating their assessment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for the AC

Summary for the Area Chair

We thank the AC and reviewers for their time. We summarize the reviews, the discussion, and the revisions.

Reviewer assessments (initial reviews)

ReviewerPresentationInformativenessSoundnessRatingEngagement
st2p3 (clear)12 (minor or fixable)1Engaged; Q1–Q3 resolved
GbiZ3 (clear)2 (learned something useful)2 (minor or fixable)2No response to rebuttal
LG5H3 (clear)2 (learned something useful)2 (minor or fixable)2"My concerns have been mostly resolved"

All three reviewers rated the presentation as clear and the soundness concerns as minor or fixable. In their initial reviews:

  • st2p: "The paper covers a clear gap in prior work as it extends distributional unlearning to the case of general data distributions. The exposition is clear and easy to follow. All notations are defined precisely and used consistently. The motivation, contribution, and treatment are highlighted sufficiently."
  • GbiZ: "Useful problem setting. Selecting a small set of important samples for unlearning is practical. It may reduce the cost of removing unwanted data." and "Good empirical results. MAMUSHI performs well across three LLMs and two datasets. The authors also test it with existing unlearning methods, such as NegGrad+ and SalUn."
  • LG5H: "The paper is mostly well-written and easy to follow. The solution, theoretical assumptions and results, and differences with existing work are clearly explained."

Discussion outcomes

  • LG5H stated that their concerns were mostly resolved. We addressed their remaining question on low-separation regimes with a theoretical explanation, a synthetic separability study, and a limitation added to the conclusion.

  • GbiZ did not respond to the rebuttal. We addressed each of their points with new results (listed below).

  • st2p stated that "Q1-Q3 are addressed" and that our clarifications on W1 are "partially fair, particularly the comments on how MAMUSHI differs from unsupervised domain adaptation." While their overall rating is a clear rejection, they rated the paper positively on presentation (clear) and soundness (minor or fixable concerns), as shown above. Their remaining objection concerns framing and positioning: they view the work as decontamination rather than unlearning. We responded that:

    1. Decontamination recovers base distributions from unlabeled mixtures, where identifiability is central, whereas our forget samples are identified, so the problem arises only under a deletion budget imposed by the cost of unlearning.
    2. Removing data and retraining is the reference notion of exact unlearning (Cao & Yang, 2015; Ginart et al., 2019; Bourtoule et al., 2021).
    3. Choosing which forget data to act on is studied within unlearning (Goel et al., TMLR 2024).

    We follow the published formulation of distributional unlearning (Allouah et al., ICLR 2026) and credit the decontamination literature in the revision.

Revisions

  • Model-level guarantee (Prop. B.9): links the selected set to the retrained model's forget and retain log-loss, explaining the budget curves.
  • Finite-sample results: a sample-size-dependent guarantee for ReLU-network classifiers (Remark B.8) and a single-error version of the transfer theorem, suggested by st2p.
  • Table 1 with standard deviations over seeds:
    • Retraining: significant improvement (21.9 ± 4.4% vs. ~59% for both baselines).
    • SalUn: parity with Coreset.
    • NegGrad+: MAMUSHI reaches the target on all three seeds (19.7 ± 6.9%), while neither baseline does within the swept budgets. Coreset runs that complete the full unlearning schedule still fall short, so the difference stems from the selected samples rather than training length; held-out seeds show the same pattern.
  • Additional baselines and ablations: non-parametric density-ratio estimators (KNN-Ratio, RuLSIF), classifier and separability ablations.
  • Positioning: attribution of classical results (Lemma 4.1, the Neyman–Pearson connection), a related-work discussion of decontamination, mixture proportion estimation and domain adaptation, an explicit statement that the selection rule is representation-agnostic, and the deletion-budget motivation in the introduction.

We hope the AC will weigh the full discussion and the revised manuscript.

Thank You.

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.