- 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
- 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
- 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.
- 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
- 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?
- 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.
- 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
Reviewer LG5H ( Rebuttal II)
i) Graceful degradation with a safety floor: as the learned logits compress toward zero (spread , ), and at exact identity the scorer collapses to chance (AUC ) with of logits in — 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- regime our Corollary~5.7 identifies as the hard case where accurate matters most.
(iii) Our real experiments operate in the healthy intermediate band: Jigsaw yields , 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., 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 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 at β=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 ( 6 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 3.46 to 4.91 (a ≈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 (P1) and retain (P2) 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 P1–P2 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 (t, df$=4$), anisotropic features in R50, with P1 shifted from P2 by a controllable amount (sep$=0$ means identical), scored by the same cross-fitted MLP used in the paper.