acceptodds
Under review as a conference paper at ICLR 2027

Overshooting the Oracle: Why Zero Forget Accuracy Is the Wrong Target for Entangled Unlearning

Abstract

Unlearning methods for entangled forget sets, where forgotten and retained data share a label, are typically trained to drive forget-set accuracy to zero. However, a model retrained without the forget set (usually considered the reference in unlearning) still retains substantial accuracy on it, given similar examples are present in the retain set. Consequently, these methods that achieve zero accuracy on the forget set do not approximate retraining; they overshoot it, at a cost to both utility and privacy. We present a separability ceiling showing that the zero-accuracy objective is infeasible precisely where the forget and adjacent retain sets are hardest to tell apart. We further argue that class-based unlearning (where the goal is to unlearn a particular subclass, e.g., the aquarium fish subclass of the superclass fish), the standard benchmark, barely tests entanglement. We then propose Swift, which replaces both endpoints of the usual objective with proxies that need no retrained model: the forget-loss distribution is matched to a held-out calibration set, and the retain-loss distribution is matched to a cheap probe fitted on retain data alone rather than to the contaminated original model. Both terms are squared 2-Wasserstein distances, and the forget term penalizes over-forgetting as much as under-forgetting. We evaluate on four datasets spanning image classification (CIFAR-100), face recognition (VGGFace2), and toxicity and social-bias detection in text (ToxiGen, SBIC). On CIFAR-100, Swift cuts the L1 deviation from the retrained reference from 134.62 to 22.37 and the membership-inference deviation from 0.556 to 0.032 compared with Two-Stage, while running up to 64× faster than retraining.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Machine unlearning, forget-retain entanglement

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_jn5XOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies machine unlearning under retain-forget entanglement and challenges the common objective of driving accuracy on the forget set toward zero. The central observation is that a model retrained from scratch without the forget set can nevertheless retain high accuracy on those forgotten examples when similar examples remain in the retain set. Therefore, aggressively reducing forget-set accuracy can “overshoot” the retrained reference rather than approximate it.

The paper formalizes this intuition through a total-variation-based separability bound relating forget accuracy and accuracy on the adjacent retain distribution. It then introduces SWIFT, which replaces the unavailable retrained model with two proxy references. The forget-set loss distribution is matched to that of held-out calibration samples from the same distribution, while retention is constrained by a one-sided Wasserstein loss relative to either a lightweight probe trained on retain data or, in some settings, the original model. The method is evaluated on CIFAR-100, VGGFace2, ToxiGen, and SBIC in class-based and instance-based unlearning settings.

Overall, I find the central critique of zero-forget-accuracy objectives useful and timely, and the instance-based formulation provides a meaningful stress test for entangled unlearning.

Critical strengths
  1. The paper identifies an important evaluation issue in approximate unlearning. The distinction between suppressing performance on the forget set and matching the behavior of a retrained model is conceptually important. The empirical examples make the issue concrete: for example, the retrained model remains highly accurate on forgotten instances when related retained samples remain.

  2. The instance-based entanglement setting is well motivated. Randomly splitting examples from the same subgroup between forget and adjacent-retain sets creates a more challenging setting than removing an entire subclass. Proposition 1 provides a clean population-level explanation of why selectively suppressing one distribution becomes impossible as the forget and adjacent distributions become indistinguishable.

  3. The experiments cover several modalities and datasets. The paper includes image classification, face recognition, toxicity classification, and social-bias/offense classification rather than demonstrating the phenomenon on a single benchmark.

  4. The paper evaluates against actual retraining. This is particularly valuable because the paper's primary claim concerns deviation from the retrained reference rather than simply minimizing forget accuracy.

  5. The paper is reasonably transparent about some limitations. In particular, it explicitly acknowledges that SWIFT can also damage held-out examples from the forget distribution; on VGGFace2 and CIFAR-100, the test-set forget-distribution accuracy remains noticeably below that of retraining.

Critical weaknesses
  1. The novelty relative to recent reference-aligned unlearning methods is unclear. The central premise, rather than maximizing the forget loss, align forgotten samples with the behavior of unseen/reference data has recently appeared in closely related work. In particular, RAUL / Multi-Objective Reference-Aligned Machine Unlearning, published in PMLR in May 2026, replaces unbounded forgetting with alignment toward a reference distribution representing unseen data, including an empirical held-out reference set, and reports closeness to retraining. Proceedings of Machine Learning Research ReGUn similarly argues for aligning forget examples to truly unseen examples via a held-out class-conditioned reference. The manuscript cites these works in the appendix, but does not experimentally compare against RAUL. The distinction claimed by the authors evaluating the calibration reference under the current parameters and introducing an additional retain-side reference is potentially meaningful, but I do not currently see sufficient evidence demonstrating that these changes constitute a substantial advance over the existing reference-alignment approach. This is especially important for RAUL because it is a peer-reviewed 2026 publication rather than merely contemporaneous unpublished work.

  2. There are serious reproducibility inconsistencies in the submitted manuscript. The main text states that the VGGFace2 experiment uses InceptionResnetV1, whereas Appendix A.6 states that VGGFace2 uses a ResNet-50, and the following page again discusses the VGGFace2 model as ResNet-50. Furthermore, Table 12 contains an unresolved placeholder, “XXX”, for the CIFAR-100 gradient-ascent learning rate. These issues make it difficult to reproduce or fully evaluate the baseline comparisons and should definitely be resolved in the rebuttal/revision.

  3. Statistical uncertainty is not reported. The main tables report individual point estimates without means, standard deviations, or confidence intervals across multiple runs. Given that the claimed advantages sometimes involve relatively small deviations from retraining, it is difficult to determine how much of the difference is robust to initialization, data splitting, or unlearning optimization. At minimum, I would like to see repeated runs for SWIFT, retraining, and the strongest relevant baselines on one representative vision and one text dataset.

Questions for authors

See weakness

Reviewer_Zm9xOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper argues that minimizing forget-set accuracy is an inappropriate objective for machine unlearning under retain–forget entanglement. When similar examples remain in the retain set, a retrained model can still correctly classify forgotten examples. Consequently, aggressively suppressing forget accuracy may overshoot the retrained reference, harming retained data and potentially increasing privacy risks. The authors establish a separability ceiling showing that the accuracy gap between forget and adjacent retain distributions is bounded by their total variation distance. They further argue that conventional class-based benchmarks insufficiently capture entanglement and introduce instance-based settings where both sets follow the same distribution. To approximate retraining efficiently, the proposed SWIFT method matches forget-set losses to held-out calibration losses under the current model and constrains retain-set losses using a probe-based or original-model reference. Experiments across four datasets show that SWIFT generally approaches retrained-model performance more closely than existing suppression-based methods in instance-based settings, with substantially lower computational cost.

Critical strengths

A more appropriate evaluation of retain–forget entanglement: The paper highlights an important limitation of conventional class-based unlearning benchmarks, where deleting an entire subclass can make the forget and adjacent retain distributions sufficiently distinguishable that reducing forget accuracy incurs little adjacent performance loss.By instead removing only a subset of instances from the same distribution, the authors demonstrate that aggressive forgetting can substantially degrade adjacent retained examples. The observation that retraining still achieves high forget-set accuracy, together with the theoretical separability ceiling, provides convincing motivation for evaluating unlearning against retrained-model behavior rather than zero forget accuracy.

Critical weaknesses

Insufficient empirical characterization of retain–forget entanglement: Although Proposition 1 relates the achievable accuracy gap to the total variation distance between the forget and adjacent retain distributions, the experiments do not systematically measure or vary this distance. They primarily compare class-based and instance-based settings, with VGGFace2 providing an additional case involving related but non-identical distributions. However, they do not systematically quantify intermediate levels of distributional similarity. Consequently, it remains unclear whether adjacent performance degradation varies continuously with distributional similarity or primarily reflects the difference between removing an entire subclass and removing only part of it.

The authors could construct intermediate entanglement levels and estimate distributional distinguishability using a two-sample classifier in a fixed feature space. Examining its relationship with adjacent performance degradation and deviation from retraining would better support the theoretical claims. Varying the forget-set fraction independently and reporting variance across random splits would further strengthen the analysis.

Limited characterization of calibration-reference reliability: The theoretical justification for the calibration reference applies to the retrained model, whereas SWIFT evaluates calibration losses under the current model. As acknowledged in Section 5 and Appendix A.9, this moving reference can degrade alongside held-out samples, allowing excessive forgetting. However, the paper does not systematically establish when this approximation becomes unreliable. For example, SWIFT over-forgets in the class-based ToxiGen setting, while its held-out accuracy on CIFAR-100 instance-based unlearning is substantially below retraining. Although the authors acknowledge these limitations, the lack of a diagnostic or sensitivity analysis makes it difficult to determine when the calibration reference can reliably approximate retrained-model behavior, particularly for deletion requests with unknown degrees of entanglement.

Questions for authors

Could the authors quantify forget–adjacent distributional similarity using an empirical proxy, such as two-sample classification accuracy, and examine its relationship with adjacent performance degradation and deviation from retraining? This would clarify how the theoretical bound relates to observed unlearning behavior across varying degrees of entanglement.

Is there a measurable criterion for identifying when the current-model calibration reference becomes unreliable? Since practical deletion requests may exhibit varying degrees of entanglement, such a diagnostic could clarify when SWIFT is likely to over-forget or under-forget.

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

This paper argues that minimizing forget-set accuracy is an inappropriate objective for machine unlearning when forget and retain samples are entangled, since a retrained model may still correctly classify forgotten samples. The authors introduce a total-variation-based separability bound and propose SWIFT, which matches forget-set losses to a held-out calibration distribution while constraining retain-set losses using a lightweight reference model. Experiments across vision and text datasets demonstrate improved alignment with retraining compared with several suppression-based methods, at substantially lower computational cost.

Critical strengths
  1. The paper clearly explains why zero-forget accuracy can lead to over-forgetting and why retraining should serve as the reference. The distinction between class-based and instance-based unlearning is particularly useful.

  2. Practical and efficient approach. SWIFT offers a relatively simple reference-based objective and shows promising results across multiple datasets, with substantial computational savings compared with full retraining.

Critical weaknesses

ReGUn (Mirlach et al., 2026) and RAUL (Khosrowshahli et al., 2026) already explore reference-guided unlearning rather than aggressive loss maximization. Although the paper discusses these methods, it does not experimentally compare against them. This makes it difficult to establish the specific advantage of SWIFT beyond existing reference-aligned approaches.

Incomplete approximation of retraining behavior. Despite improvements over suppression-based methods, SWIFT still substantially underperforms retraining on held-out samples from the forget distribution (70% vs. 87% on CIFAR-100; 79.34% vs. 90.61% on VGGFace2). Moreover, matching marginal loss distributions does not guarantee behavioral equivalence to retraining. These limitations weaken the paper's central effectiveness claim.

Questions for authors

Can the authors provide direct comparisons with ReGUn and RAUL under the same experimental settings, and clarify which components of SWIFT offer advantages over these methods?

Why does SWIFT still significantly degrade accuracy on unseen samples from the forget distribution? Can the authors provide stronger evidence of behavioral equivalence to retraining, beyond aggregate accuracy and loss-based membership inference?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.