acceptodds
Under review as a conference paper at ICLR 2027

Forgotten Compared to What? A Calibrated Audit of Class Unlearning

52 views

Abstract

Machine unlearning aims to make a trained model behave like one retrained without designated data. Whether it succeeds depends on what we measure. A model may stop predicting a removed class even though its scores still distinguish that class. Yet finding such a signal is not enough: a model trained without the class may also recognize its images through generalization. The relevant question is therefore: forgotten compared to what? We introduce a calibrated audit that measures recovery from an unlearned model relative to models trained without the removed class, at a fixed operating point. Our main test, woken recall, measures recall after calibrating a single logit offset on retained data at a nominal 5% false-positive rate. We use the existing classifier and choose the offset using only retained-class examples. In a CIFAR-10 study of 480 checkpoints from 16 fixed unlearning configurations, models with near-zero forget accuracy often have higher removed-class recall than retain-only references matched on retained accuracy. A frozen-feature probe gives a separate, more supervised view and leads to the same conclusion. Recovery has to be read against a stated reference and operating point; forget accuracy alone cannot provide that comparison.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Machine Unlearning, Computer Vision

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_1vxbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper discusses the class-wise unlearning evaluation. Existing evaluation focuses on comparing classification accuracies with a retrained reference model, while authors claim that many unlearned methods still remain more forget-class recoverability than a totally retrained model. Authors propose a method from the logits layer to calibrate forget-class logits by using only remain-class data for auditing forget-class recoverability.

Critical strengths
  1. The paper emphasizes that recoverability should be evaluated relative to retain-only models. The motivation is clear, reasonable and of significance for unlearning topic.
  2. The study evaluates 16 unlearning configurations across multiple classes and seeds. The authors further examine utility-matched versus fully trained references, BatchNorm recalibration, and operating-point sensitivity. These analyses provide convincing evidence that excess recovery is widespread in the evaluated CIFAR-10 setting.
Critical weaknesses
  1. The authors claim that an unlearned model exhibits greater recoverability than a retrained model. However, their discussion focuses primarily on the logits layer, with relatively little analysis or exploration of differences in feature distributions.

  2. The retain-only references preserve a 10-way classification head despite receiving no positive training examples for the removed class. Consequently, the removed-class head is trained only through negative supervision, which may produce different score distributions from those of a standard 9-way retrained model. Since Woken Recall directly relies on this head, the observed excess recovery may partly reflect differences in classifier-head training rather than residual class information.

  3. The audit distinguishes removed-class examples from retained classes but does not establish whether the recovered signal is specific to the removed-class or reflects a more general response to complete noise or OOD inputs. Comparisons with 9-way retrained backbones using feature probes, together with OOD controls, would help clarify the interpretation of the observed recovery gaps.

Questions for authors

Please refer to the above.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 1vxb (1/2)

Thank you for the constructive review and for the specific controls you suggested. We ran both of them, and have reported the findings below. Gaps are medians in percentage points. 95% bootstrap CIs resample removed classes or (method, class) cells, as in Appendix D.3.

W1. Feature-level analysis

Comment. The discussion focuses on the logits layer, with relatively little analysis of feature distributions.

Answer. The discussion covers both the logits layer and the feature distributions. The paper already has a feature-level readout, the frozen-feature probe in Section 3 (“Supervised feature recovery”), which may be easy to miss there. To further expand the analysis of feature distributions, this response adds three new analyses (listed at the end of this answer). The probe fits a logistic classifier on the frozen features of each model. It is compared with the same matched retain-only references as woken recall. All main results report both readouts.

The two readouts answer different questions. Woken recall asks what the released head still encodes, which one scalar can restore. The probe asks what a newly fitted linear classifier can extract from the features, as in [1]. Each is compared with its own matched reference, so together they can better show where the excess sits.

CIFAR-100 method (saved)Woken gapProbe gapWhere the excess sits
CF-k+4.0+19.5Features
SalUn+28.0+5.5Head

CF-k fine-tunes the last layers on retained data. Its head resembles the reference, while its features keep the class. SalUn shows the reverse.

W2 compares the features of 9-way and 10-way retain-only models. W3 tests whether the feature-level signal is specific to the removed class. In multi-class forgetting, the probe gap is positive in all 120 high-utility rows (median +11.05, 95% CI [8.35, 12.7]).

Proposed text (Section 4.6, after "Woken recall reads existing classifier scores; a feature probe asks a different question after fitting a new readout.").

Together, the two readouts show where the excess sits. On CIFAR-100, CF-k keeps its excess in the features and SalUn keeps it in the head (new table in Appendix H).

W2. The 10-way head of the retain-only references

Comment. The retain-only references keep a 10-way head, so the removed-class row is trained only through negative supervision. Its scores may differ from those of a standard 9-way retrained model. Since woken recall relies on this head, the excess may reflect how the head was trained.

Answer. The excess does not reflect how the head was trained. The matched retain-only references keep the same 10-way head. So the removed-class row gets the same negative supervision in the unlearned model and in its reference. We kept the 10-way head for two reasons.

  1. Unlearned models keep the 10-way head. The matched comparison for a readout of that head is a retain-only model with the same head. A 9-way model has no score for the removed class, so woken recall cannot be computed on it.
  2. Negative-only supervision of the removed-class row is the state of a model that never saw the class. An unlearned model that truly matches retraining should reproduce it. This low floor is also what gives woken recall room to separate methods.

We also report results with 9-way retain-only references, as you suggest. The probe does not use the head, so 9-way references test it directly. We trained 9-way retain-only models for all 10 classes with 3 seeds each, using the same recipe and the same checkpoint epochs as the 10-way references. This gives 30 new training runs and 870 checkpoints. Fully trained 9-way and 10-way references reach the same retain accuracy (95.1% for both).

ComparisonSavedBN recalibrated
9-way minus 10-way probe recall, at matched retain accuracy+0.2 [0.0, 0.5]+0.2 [-0.2, 0.6]
Same, fully trained references+0.4 [-0.4, 0.8]+0.5 [-0.6, 1.8]
High-utility probe gap, 10-way to 9-way references10.70 to 10.4514.00 to 14.20
Sign changes among high-utility rows0 of 2640 of 189

The head choice leaves the features unchanged. Against 9-way references, the probe gaps stay the same, and no high-utility row changes sign. Across all saved rows with both references, 2 of 362 change sign, and both are UNSIR.

Proposed text (Section 3, "Models and fixed configurations", after the sentence on the removed output).

We keep the ten-way head because unlearned models keep it, and a nine-way model has no removed-class score for woken recall. Nine-way retain-only models give the same probe recall at matched retain accuracy (new Appendix D.7).

AuthorsOpenReview ↗

Response to Reviewer 1vxb (2/2)

W3. Specificity to the removed class

Comment. The audit does not establish whether the recovered signal is specific to the removed class, or a general response to noise and OOD inputs. Comparisons with 9-way backbones and OOD controls would help.

Answer. The recovered signal is specific to the removed class. Unlearned models show no excess response to noise or OOD inputs. We establish this with new OOD controls, Gaussian noise inputs, and the 9-way backbones in W2. The first OOD set is SVHN [2]. The second is the CIFAR-100 classes that do not overlap with any CIFAR-10 class (5,500 images from 11 superclasses).

If the excess were a general response to unfamiliar inputs, unlearned models would also respond more to OOD images. We measured how well each model's score ranks removed-class images above OOD images (AUROC). We also measured recall at a threshold that holds the average false-positive rate on retained and OOD images at 5%.

High-utility, saved, matched referencesWoken gap (positive of 264)Probe gap (positive of 264)
AUROC, removed class vs SVHN+48.3 [44.8, 53.7] (260)+4.7 [3.0, 8.0] (228)
AUROC, removed class vs CIFAR-100 OOD+48.8 [40.3, 53.1] (264)+7.7 [6.5, 9.1] (262)
Recall at OOD-aware threshold, SVHN+56.1 [46.5, 74.4] (256)+19.3 [12.9, 24.1] (248)
Recall at OOD-aware threshold, CIFAR-100 OOD+56.55 [44.6, 76.4] (261)+28.35 [24.0, 32.4] (263)

Unlearned models rank removed-class images above OOD images far more reliably than matched references do, on both readouts. BN-recalibrated models give the same picture. Gaussian noise also gives no general response. It rarely crosses either threshold on unlearned models (median 3.0% for woken recall, 0.4% for the probe).

One more result supports the paper's main point. On SVHN, the calibrated woken threshold fires on about half of the images, for unlearned models (48.9%) and matched references (49.6%) alike. An absolute firing rate therefore says little. Recovery has to be read against a reference.

Proposed text (new Appendix D.8, with one sentence in Section 4.1).

We test class specificity with two OOD sets, SVHN and the CIFAR-100 classes with no CIFAR-10 counterpart. On both readouts, unlearned models rank removed-class images above OOD images more reliably than matched references (Appendix D.8). The excess is specific to the removed class.

We hope this clarification, together with the supporting evidence and corresponding manuscript revisions, satisfactorily addresses all aspects of your concern.

References

  1. Y. Gao, A. Unal, A. Rangamani, Z. Zhu. An Illusion of Unlearning? Assessing Machine Unlearning Through Internal Representations. AISTATS 2026. https://arxiv.org/abs/2604.08271
  2. Y. Netzer, T. Wang, A. Coates, A. Bissacco, B. Wu, A. Y. Ng. Reading Digits in Natural Images with Unsupervised Feature Learning. NIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011.
AuthorsOpenReview ↗

Gentle reminder for response: revised manuscript uploaded

Dear Reviewer,

The discussion period ends soon, so we wanted to follow up on our response. We ran both controls you suggested. The revised manuscript is uploaded, with all changes in blue. Each of your concerns has been addressed in the paper:

  1. Feature-level analysis: Section 4.6 and Appendix H
  2. Nine-way retain-only references: Appendix D.7
  3. OOD and noise controls for class specificity: Appendix D.8

We would be grateful to hear whether this resolves your concerns. If it does, we hope you will consider updating your score. We are happy to answer any further questions.

Sincerely, The Authors

Reviewer_BPmaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a calibrated framework for evaluating class unlearning using woken recall and linear probes. Experiments on different datasets, model architectures, and unlearning methods show that information about forget classes can still be recovered, even when standard metrics like forget accuracy suggest successful forgetting.

Critical strengths
  1. The paper addresses an important limitation of using forget accuracy alone to evaluate machine unlearning. Comparing the information recovered from unlearned models with models retrained without the forget class provides a more informative way to assess unlearning effectiveness.

  2. The authors evaluate several unlearning methods on CIFAR-10 and provide additional experiments on CIFAR-100 and ViT models. They also consider both logit-based recovery and representation-based recovery, offering complementary perspectives on residual class information.

Critical weaknesses

1.Limited novelty compared with existing unlearning evaluations. Previous studies have already shown that low forget accuracy does not necessarily indicate successful removal of class information and that residual information can be recovered through classifier adjustments or representation-level probes. Although the proposed reference-relative calibration is useful, the paper could better distinguish its methodological contribution from existing evaluation approaches, particularly those using linear probes and retraining comparisons.

2.Limited experimental settings and uneven evaluation across methods. Most experiments focus on single-class unlearning on CIFAR-10 with ResNet-18, with limited evaluation on other datasets and architectures. The paper also does not consider simultaneous multi-class forgetting. In addition, some evaluated methods show poor forgetting–utility trade-offs, while others have very few reference-supported checkpoints, making it difficult to draw equally strong conclusions across all evaluated methods. These limitations raise questions about how well the proposed framework generalizes to more complex unlearning settings and larger models.

Questions for authors
  1. How sensitive is the proposed audit to the choice and number of forget classes? Would the calibration procedure still work if multiple forget classes were considered simultaneously?

  2. How sensitive are the main woken-recall and probe gaps to different calibration and evaluation partitions? Could the authors report repeated-split variability to establish whether the observed differences are robust?

  3. Given the limited number of reference-supported checkpoints for some methods, how confident can we be that the reported recovery trends generalize across different unlearning approaches? Could the authors provide additional analysis to clarify this?

5 comments
AuthorsOpenReview ↗

Response to Reviewer BPma (1/4)

Thank you for the careful review and the concrete questions. We ran new analyses for each of them, and have reported the new results below. Gaps are medians in percentage points. 95% bootstrap CIs resample removed classes or (method, class) cells, as in Appendix D.3.

W1. Novelty against probe and retraining comparisons

Comment. Earlier work already shows that low forget accuracy does not mean the class is gone, and that residual information can be recovered by adjusting the classifier or probing representations. The paper should separate its method from evaluations that use linear probes and retraining comparisons.

Answer. The paper does not claim that residual information can be recovered. Its title, “Forgotten Compared to What?”, states the central claim. Recovery after unlearning has meaning only against a reference, and the paper builds that reference. We cite the earlier work that recovers residual information [1, 2, 3]. Comparing against a retrained model is standard practice. Our contribution is how that comparison is designed. We separate our method from linear probes and retraining comparisons in two ways.

1. Retain-only references matched in utility. Standard retraining comparisons use one fully trained retrain. A retain-only model recovers the removed class better as it trains. So a model that loses utility recovers the class less, even when nothing class-specific was removed, because of feature overlap.

Against one fully trained reference, that damage can falsely read as forgetting. Matching at equal retain accuracy removes this confound.

Method (CIFAR-10, saved)Retain accuracyProbe gap vs fully trained referenceProbe gap vs matched reference
GradientAscent77%-10.8 [-13.2, -8.1]+9.0 [-4.7, 17.4]
UNSIR76%-18.1 [-25.0, -11.1]-4.5 [-7.7, 1.4]

Against the fully trained reference, both methods appear to remove more than retraining does. At equal utility, that advantage is gone.

2. Woken recall reads the released head. A linear probe fits a new classifier on frozen features. Woken recall uses the released classifier and one scalar calibrated on retained data. It ignores a constant shift of the removed-class logit, which is the cheapest way to reach 0% forget accuracy [3].

Its retain-only floor also stays low, because a retain-only head never sees a positive example of the removed class. Features transfer, so probe floors are high. On ImageNet-pretrained ViT-S/16, retain-only references reach about 99.8% probe recall, which leaves the probe 0.2 points to separate methods.

ViT-S/16 methodWoken gapProbe gap
SCRUB-43.4-0.6
CF-k, FineTune, SalUnabout +51+0.2

Where both readouts have room, as on high-utility CIFAR-10 checkpoints, they agree. Two readouts built this differently reach the same answer, which is independent evidence for the effect.

Proposed text (Section 5, extending the sentence on Gao et al.).

Gao et al. [1] and Kim et al. [2] show that unlearned models keep class information in their features, and [1] observes the same for retain-only models. Our audit differs in three ways. We compare against retain-only references matched in retain accuracy, because utility loss otherwise reads as forgetting. Woken recall reads the released head, whose retain-only floor stays low when features transfer, as on our ViT models. It also ignores a shift of the removed-class logit, so it measures what one scalar restores.

We believe this clarification directly addresses the concern regarding how our method is separated from linear probes and retraining comparisons.

AuthorsOpenReview ↗

Response to Reviewer BPma (2/4)

W2 and Q1. Multi-class forgetting and the choice of forget class

Comment. Most experiments forget one CIFAR-10 class, and simultaneous multi-class forgetting is missing. Q1 asks how sensitive the audit is to which classes and how many are forgotten, and whether the calibration still works with several forget classes.

Answer. Yes, the calibration still works with several forget classes. Which classes and how many are forgotten change the size of the gap, and the matched reference accounts for this. The sign of the gap holds in every forget set. Each removed class gets its own offset, calibrated on retained-class images only. Recall is then measured on that class's held-out images. The matched references are retain-only models trained without the whole forget set.

We ran three forget sets with 3 seeds and 14 methods, and trained new retain-only reference ladders for each set. Every high-utility row is positive on both readouts, in every set.

Forget set (saved, high-utility)Rows positiveWoken median gapProbe median gapReference woken recall
cat + dog36 / 36+32.35+12.548.2%
airplane + frog30 / 30+45.0+10.435.2%
automobile + bird + ship54 / 54+47.6+7.8523.0%
Pooled120 / 120+41.8 [33.1, 46.2]+11.05 [8.35, 12.7]

In the BN-recalibrated variant, all 82 high-utility rows are positive, with medians of +37.25 (woken) and +12.3 (probe).

The choice of class matters in the way the reference predicts. Cat and dog resemble each other, so a retain-only model already recovers the removed one more often, and the gap is smaller. The matched reference absorbs this, so classes of different difficulty stay comparable. In the main study, every one of the ten CIFAR-10 classes is a target, and our CIs account for variation across classes (Appendix D.3).

We kept each method's single-class configuration. With several classes, three methods left the high-utility cohort. SCRUB lost most of its retain accuracy, SSD stopped forgetting, and BadTeacher failed to forget frog in the airplane and frog set. POUR-D and POUR-P project out a single class vector, so we did not run them.

Proposed text (new Appendix I, with one sentence added to the Scope paragraph of Section 4.6).

We extend the audit to multi-class forgetting with three forget sets of two and three classes. Each removed class gets its own offset calibrated on retained classes, and references are retain-only models trained without the whole set. All 120 high-utility rows are positive on both readouts. The gap is smallest for cat and dog, where the reference already recovers the removed class most often.

W2. Other datasets and larger models

Comment. Evaluation beyond CIFAR-10 and ResNet-18 is limited, which raises questions about larger models.

Answer. The evaluation goes beyond CIFAR-10 and ResNet-18. The submission covers CIFAR-100 and an ImageNet-pretrained ViT-S/16, where it reported only the probe. We have now computed woken recall on all 420 checkpoints of these extensions. These are 66 unlearned models and 354 retain-only reference checkpoints. Woken recall now covers three settings instead of one, across two datasets and two architectures.

DatasetVariantWoken: positive / supportedWoken median gapProbe: positive / supportedProbe median gap
CIFAR-100Saved25 / 26+26.023 / 26+14.5
CIFAR-100BN recalibrated30 / 30+20.527 / 30+12.0
ViT-S/16Saved27 / 36+32.419 / 36+0.2

ViT-S/16 is ImageNet-pretrained, the setting where feature probes have the least room. This pretrained setting is the one closest to larger models, and woken recall keeps its resolution there. Further scaling is limited by compute constraints.

Proposed text (Section 3, "Reference support and analysis history"). This replaces the sentence that assigns the CIFAR-100 and ViT checks to the probe.

Woken gaps use matched checkpoints in all three settings, while endpoint and trajectory checks use the probe.

Proposed text (opening of Appendices G and H). This replaces the sentences that limit both extensions to the probe.

We report woken recall next to the probe. On ViT, where the probe saturates, woken recall separates SCRUB (-43.4) from CF-k, FineTune and SalUn (about +51).

AuthorsOpenReview ↗

Response to Reviewer BPma (3/4)

Q2. Variability across calibration and evaluation splits

How sensitive are the woken-recall and probe gaps to the calibration and evaluation partitions? Please report repeated-split variability.

Answer. The woken-recall and probe gaps are not sensitive to the calibration and evaluation partitions. We repeated the full analysis on 20 random calibration and evaluation splits. For each split we recomputed the offsets, the probes and every gap.

High-utility cohortOriginal splitMedian of 20 splitsIQR across splitsPositive targets across splits
Saved, woken+52.20+51.6851.11 to 52.51254 to 257 of 264
Saved, probe+10.70+10.439.98 to 10.61262 to 264 of 264
BN recalibrated, woken+44.00+44.2543.78 to 44.73181 to 186 of 189
BN recalibrated, probe+14.00+12.9512.60 to 13.25187 to 189 of 189

The IQR of the median gap is at most 1.4 points, and the count of positive targets moves by at most 5. Of the 264 saved high-utility targets, 254 are positive in all 20 splits on woken recall and 259 on the probe. All conclusions hold.

We also selected matched references on a held-out half of the test set, as asked by Reviewer 7Df5. Median gaps moved by at most 0.25 points.

Proposed text (Appendix D.4). This replaces the paragraph beginning "Repeated calibration and probe refitting", and the matching clause in the Scope paragraph of Section 4.6 is removed.

We repeat the analysis on 20 random calibration and evaluation splits. A new table reports the median gap, its IQR across splits and the range of positive counts. All conclusions hold on every split.

We hope this clarification addresses the concern regarding repeated-split variability.

W2 and Q3. Uneven support across methods

Comment. Some methods have poor forgetting-utility trade-offs, and some have few reference-supported checkpoints. Q3 asks how confident we can be that the recovery trends hold across methods.

Answer. Ten methods have 30 of 30 targets reference-supported, and SCRUB has 28 of 30 (93%). So 11 of 16 methods have near-full support. The recovery trends hold across the ten fully supported methods, with tight CIs. We computed support counts and 95% CIs for every method, clustered by class.

Method (CIFAR-10, saved, all targets)Supported of 30Woken gap [95% CI]Probe gap [95% CI]
BadTeacher30+66.4 [55.4, 75.1]+12.3 [7.7, 19.5]
POUR-D30+65.8 [57.0, 73.9]+11.7 [7.4, 17.9]
POUR-P30+63.7 [53.2, 71.4]+12.0 [7.5, 17.9]
SSD30+63.3 [54.0, 68.9]+11.8 [7.3, 17.6]
Spotter30+61.6 [50.4, 67.1]+10.5 [6.0, 15.3]
NegGradPlus30+50.7 [45.8, 59.4]+11.5 [6.4, 15.8]
InfluenceUnlearning30+47.7 [36.3, 59.8]+9.8 [3.7, 12.4]
FineTune30+37.7 [32.1, 45.7]+11.0 [5.7, 15.8]
SalUn30+37.5 [31.3, 53.0]+11.2 [6.9, 16.7]
L1Sparse30+36.0 [29.3, 41.4]+11.0 [6.6, 17.2]
SCRUB28+8.3 [-3.0, 24.8]+7.2 [3.0, 9.8]
UNSIR21+2.3 [-3.6, 10.2]-4.5 [-7.7, 1.4]
GradientAscent20+4.8 [0.3, 7.7]+9.0 [-4.7, 17.4]

Ten methods show excess on both readouts, with CIs above zero. SCRUB and UNSIR are consistent with their references on woken recall, and GradientAscent shows a small excess. So the audit separates methods, and gives a positive verdict only where the data support one.

Three methods have little or no support. BoundaryExpand and FisherForgetting have none, and BoundaryShrink has 3 of 30. Their retain accuracy (11% to 30%) is too low to find at least five retain-only checkpoints within 2 points. For these, the audit has no matched reference and does not give any verdict.

On poor trade-offs, the high-utility cohort (forget accuracy at most 1%, retain accuracy at least 90%) keeps only checkpoints that forget well and keep their utility. In it, 255 of 264 targets are positive on woken recall and 263 of 264 on the probe, with medians of +52.2 [47.7, 56.8] and +10.7 [9.0, 12.8].

Proposed text (Table 8, Appendix C).

Table 8 adds per-method woken gaps and 95% CIs for both readouts, resampling removed classes. We draw per-method conclusions only where the CI excludes zero. Methods without matched support stay unsupported, as in Figure 3.

Thus, the above analysis shows that the recovery trends hold across methods.

We hope this clarification, together with the supporting evidence and corresponding manuscript revisions, satisfactorily addresses all aspects of your concern.

Reviewer_7Df5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a calibrated, reference-relative audit for evaluating class unlearning. Its main metric, woken recall, measures forgotten-class recovery using existing classifier margins at a fixed false-positive rate, without fitting a new classifier. Recovery is compared against retain-only models, including references matched by retained-class accuracy. The authors additionally evaluate supervised frozen-feature probes.

Experiments across 16 unlearning configurations and 480 CIFAR-10 checkpoints show that models achieving near-zero forget accuracy often exhibit greater recoverable class discrimination than retrained references. The paper also examines reference selection, BatchNorm recalibration, operating-point sensitivity, and smaller-scale extensions to CIFAR-100 and pretrained ViTs.

Critical strengths

The paper provides extensive experiments across 16 unlearning configurations and 480 CIFAR-10 checkpoints. The additional analyses of normalization, operating points, and reference selection strengthen confidence that the reported recovery gaps are not merely artifacts of one particular evaluation configuration.

  • Comparing recovery against retain-only references addresses an important limitation of absolute recovery metrics: models trained without a class can still recognize examples belonging to it. The paper demonstrates that reference choice and operating points can substantially change the interpretation of recovery measurements.
Critical weaknesses

Limited novelty and generalizability

Novelty: The central observation that low forget accuracy can conceal recoverable forgotten-class information has already been demonstrated in prior peer-reviewed work. In particular, Gao et al. [1] investigate feature–classifier misalignment and demonstrate recovery through linear probing despite apparently successful unlearning. Kim et al. [2] similarly establish that unlearning methods can preserve internal representations while primarily modifying classifier behavior. Against this background, the primary methodological contribution, woken recall, amounts to measuring recall at a fixed false-positive rate using existing classifier margins. Its invariance to constant class-logit shifts follows directly from standard properties of score thresholding. Although the calibrated existing-score readout and accuracy-matched reference construction are distinct contributions, their incremental value over existing representation-based evaluations is not convincingly established.

Generalizability: The systematic evaluation primarily concerns CIFAR-10/ResNet-18. The CIFAR-100 and pretrained ViT extensions are narrower and exclusively evaluate the supervised feature probe rather than the principal proposed metric. Although authors cite these papers, this still leaves the generality of woken recall across datasets and architectures insufficiently demonstrated. Furthermore, the paper reports substantial agreement between woken recall and supervised feature probes among high-utility checkpoints. Given that existing literature already establishes the effectiveness of representation-level recovery diagnostics, it remains unclear when the proposed metric yields qualitatively different or more actionable conclusions.

[1] https://proceedings.mlr.press/v300/gao26e.html [2] https://arxiv.org/abs/2503.06991

Questions for authors
  • Beyond the computational simplicity of applying a scalar offset, what meaningful conclusions does woken recall enable that established representation-level recovery evaluations, particularly those of Gao et al. (2026), cannot?
  • Why are the CIFAR-100 and pretrained ViT extensions primarily evaluated using supervised probes rather than woken recall?
  • How sensitive are the reported gaps to selecting utility-matched references using a separate validation set instead of test accuracy?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 7Df5 (1/3)

Thank you for the constructive review. We ran new analyses for each of them, and have listed the new results below. Gaps are medians in percentage points. 95% bootstrap CIs resample removed classes or (method, class) cells, as in Appendix D.3.

W1 and Q1. Novelty against [1] and [2], and what does woken recall add?

Comment. Gao et al. [1] and Kim et al. [2] already show that low forget accuracy can hide recoverable class information. Woken recall reads existing margins at a fixed false-positive rate, and its value over representation-level evaluations is not established. Q1 asks what woken recall concludes that [1] cannot.

What we take from [1] and [2]. The paper does not claim that low forget accuracy can hide recoverable class information. Its title, “Forgotten Compared to What?”, states the central claim. Recovery has meaning only against a reference, and the paper builds that reference. We cite [1] and [2] for the recovery finding. [1] also observes that a retain-only model’s features still separate the forgotten class, because features transfer. Comparison against a retain-only model is standard practice, and [2] uses one as its target.

What we add. Our contribution is how the measurement is built. Three choices set it apart, and each leads to a conclusion that differs from the evaluations in [1, 2].

We also see value in the agreement itself. Woken recall reads the released classifier with one scalar calibrated on retained data. The probe in [1] fits a new classifier on features. Two readouts built this differently agree on high-utility CIFAR-10 checkpoints, where both have room to move.

Below is what woken recall concludes that [1] cannot:

1. Matched references separate forgetting from damage. A retain-only model recovers the removed class better as it trains. On CIFAR-10, retain-only references matched to GradientAscent (about 77% retain accuracy) reach 65.2% probe recall, against 88.4% for fully trained references. A damaged model therefore recovers the class less, even when nothing class-specific was removed.

Against one fully trained reference, damage to the entire model reads as forgetting. [1] treats probe accuracy below that of retraining as a goal for future methods [1, Sec. 4.1]. We see the target differently. Unlearning should match a model that never saw the class, so falling below that model is also non-ideal. At equal utility, such a drop means the model still recognizes the class in order to suppress it. An auditor comparing with full retraining would read this as success. For example, UNSIR with the CMF head in [1, Table 3] reaches 31.2% probe forget accuracy, against 77.4% for retraining, while output retain accuracy moves only from 94.7% to 91.8%. By our matched comparison, a drop this large at nearly equal utility would count as class-specific suppression. Our matched references compare at equal retain accuracy, so they separate this case from a utility loss.

Method (CIFAR-10, saved)Retain accuracyProbe gap vs fully trained referenceProbe gap vs matched reference
GradientAscent77%-10.8 [-13.2, -8.1]+9.0 [-4.7, 17.4]
UNSIR76%-18.1 [-25.0, -11.1]-4.5 [-7.7, 1.4]

Against the fully trained reference, both methods appear to remove more than retraining does. Against a reference of equal utility, that advantage is not present.

[1] notes the same trade-off for NegGrad+, which loses some retain accuracy while its probe accuracy falls below retraining [1, Sec. 4.1]. Matching corrects for it. The correction is small above 90% retain accuracy (0.8 points, Section 4.3) and large below 85% (13.4 points), which is where methods appear to beat retraining.

2. Woken recall keeps a low floor where feature probes saturate. Features transfer, so a retain-only model's features still separate the removed class. A retain-only head never sees a positive example of that class, so it cannot rank it. The two floors differ sharply.

Retain-only reference (median)Probe recallWoken recall
CIFAR-10, ResNet-18, fully trained (Table 13)88.4%36.4%
ViT-S/16, ImageNet-pretrained, matched (Section 4.5)about 99.8%about 48%

On ViT the probe has 0.2 points of headroom and cannot separate methods. [1, Table 9] shows the same saturation, with retraining at 97.71% probe forget accuracy on CIFAR-10 and unlearned methods between 96.15% and 98.64%. Woken recall still separates methods.

ViT-S/16 methodWoken gap (positive of 9)Probe gap
SCRUB-43.4 (0)-0.6
CF-k+50.6 (9)+0.2
FineTune+51.0 (9)+0.2
SalUn+51.0 (9)+0.2

Better pre-training makes features generalize better, which pushes every feature-level floor toward its ceiling. [2] meets the same limit, with retrained and original models already 88% similar by CKA [2, Sec. 4.3], and responds with a harder forgetting scenario. Woken recall keeps its resolution in the standard scenario.

AuthorsOpenReview ↗

Response to Reviewer 7Df5 (2/3)

3. Woken recall ignores the cheapest way to suppress a class. Adding a constant to the removed-class logit does not change woken recall (Proposition 1). That shift is the shortcut that makes 0% forget accuracy cheap. Zheng et al. [3] show that changing the removed-class bias alone drives forget accuracy to near zero while retain accuracy stays high. Woken recall measures the ranking that is left, restored by a single scalar addition.

This tests a mechanism. Under neural-collapse assumptions, [1, Prop. 1] predicts that NegGrad turns the forget-class weight to point away from the class mean. Such a head scores forget-class images lowest on their own logit, so woken recall would be near zero.

On CIFAR-10, NegGradPlus reaches 78.4% woken recall, against 26.0% for matched references (median gap +50.7). In trained networks, most of the suppression is an offset that our scalar addition undoes.

The probe and woken recall also act as two attackers of different strength. Woken recall needs logits and one scalar. The probe needs features and labeled forget data. Each is compared with its own matched reference, so together they can better show where the excess exactly sits, rather than just verify the excess exists.

CIFAR-100 method (saved)Woken gapProbe gapWhere the excess sits
CF-k+4.0+19.5Features
SalUn+28.0+5.5Head

CF-k fine-tunes the last layers on retained data, so its head resembles the reference while its features keep the class. SalUn shows the reverse. A feature-level probe as in [1] would rate SalUn's residual as small.

4. Practical differences. Woken recall runs on logits alone, so an auditor can apply it to a model served behind an API (in soft-label or score-based black-box model). Each audit is just one forward pass and one quantile, with no training. Refitting the probe on the same split moved individual gaps by up to 1.2 points, while woken recall returned identical values.

The matched references cost no extra training. Checkpoints saved during one retain-only run already form the ladder (Appendix B.2), so the cost matches the single retrain that [1, 2] use.

Proposed text (Section 5, extending the sentence on Gao et al.).

Gao et al. [1] and Kim et al. [2] show that unlearned models keep class information in their features, and [1] observes the same for retain-only models. Our audit differs in three ways. We compare against retain-only references matched in retain accuracy, because utility loss otherwise reads as forgetting. Woken recall reads the released head, whose retain-only floor stays low when features transfer, as on our ViT models. It also ignores a shift of the removed-class logit, so it measures what one scalar restores.

Proposed text (Section 4.3, after the paragraph on Figure 4).

We add GradientAscent and UNSIR to this analysis. Against fully trained references, both appear to remove more than retraining (-10.8 and -18.1). Against matched references, their gaps are +9.0 and -4.5, with CIs that include zero.

Proposed text (Section 4.2).

Under neural-collapse assumptions, [1] predicts that NegGrad points the forget-class weight away from the class mean, which would give near-zero woken recall. NegGradPlus instead reaches 78.4%, against 26.0% for matched references. In trained networks, most suppression is an offset that one scalar undoes.

We believe this clarification directly addresses the concern regarding the value of woken recall alongside representation-level evaluations.

AuthorsOpenReview ↗

Response to Reviewer 7Df5 (3/3)

W1. Proposition 1 follows from thresholding

Comment. The invariance of woken recall to constant class-logit shifts follows directly from standard properties of score thresholding.

Answer. The paper states Proposition 1 as a design requirement for woken recall. It is not listed among the contributions. Forget accuracy lacks invariance to constant class-logit shifts, so one shift reaches 0% forget accuracy cheaply [3]. Woken recall is built to have this invariance, and the text now says it follows from standard properties of score thresholding.

Proposed text (Section 2.4, before Proposition 1).

The next property is standard for thresholded scores. We state it because forget accuracy lacks it.

W2 and Q2. Woken recall beyond CIFAR-10

Comment. The CIFAR-100 and ViT extensions report only the supervised probe, so the generality of woken recall is not shown.

Answer. The generality of woken recall is now shown across the full study. We have computed woken recall on all 420 CIFAR-100 and ViT-S/16 checkpoints, with the same calibration and matched references. Together with multi-class forgetting, woken recall now covers two datasets, two architectures, and single- and multi-class forgetting. It gives positive gaps in every setting, and separates methods on ViT, where the probe is unable to.

DatasetVariantWoken: positive / supportedWoken median gapProbe: positive / supportedProbe median gap
CIFAR-100Saved25 / 26+26.023 / 26+14.5
CIFAR-100BN recalibrated30 / 30+20.527 / 30+12.0
ViT-S/16Saved27 / 36+32.419 / 36+0.2

Woken recall gives positive gaps on both datasets. On ViT it is the only readout with room to separate methods, as shown per method in our answer to Q1. ViT has no BatchNorm layers, so the recalibrated variant does not apply.

In the high-utility ViT cohort only SCRUB and SalUn remain, with opposite signs. We therefore report ViT per method.

We also tested multi-class forgetting, which Reviewer BPma asked about. Across three forget sets of two and three CIFAR-10 classes, all 120 high-utility rows show a positive woken gap (median +41.8, 95% CI [33.1, 46.2]).

Proposed text (Section 3, "Reference support and analysis history"). This replaces the sentence that assigns the CIFAR-100 and ViT checks to the probe.

Woken gaps use matched checkpoints in all three settings, while endpoint and trajectory checks use the probe.

Proposed text (opening of Appendices G and H). This replaces the sentences that limit both extensions to the probe.

We report woken recall next to the probe. On ViT, where the probe saturates, woken recall separates SCRUB (-43.4) from CF-k, FineTune and SalUn (about +51).

Q3. Selecting references on held-out data

Comment. How sensitive are the gaps to selecting matched references on a separate validation set instead of test accuracy?

Answer. The gaps are not sensitive to selecting matched references on a separate validation set. To show this, we split the test set in two. We selected matched references by retain accuracy on one half (4,500 retained images) and measured every gap on the other half.

VariantReadoutSupported targetsMedian gap
SavedWoken372 to 369+49.45 to +49.70
SavedProbe372 to 369+10.25 to +10.40
BN recalibratedWoken382 to 382+55.10 to +55.20
BN recalibratedProbe382 to 382+13.00 to +13.10

Each cell reads from the current protocol to held-out selection. Medians move by at most 0.25 points. In the high-utility cohort, support and positive counts do not change at all.

Proposed text (Appendix D.6). This replaces the sentences stating that a validation-based matching check has not been performed.

We repeat the analysis with matched references selected on the calibration half and all gaps measured on the evaluation half. Median gaps change by at most 0.25 points, and the high-utility cohort is unchanged.

We hope this clarification, together with the supporting evidence and corresponding manuscript revisions, satisfactorily addresses all aspects of your concern.

References

  1. Y. Gao, A. Unal, A. Rangamani, Z. Zhu. An Illusion of Unlearning? Assessing Machine Unlearning Through Internal Representations. AISTATS 2026. https://arxiv.org/abs/2604.08271
  2. Y. Kim, S. Cha, D. Kim. Are We Truly Forgetting? A Critical Re-examination of Machine Unlearning Evaluation Protocols. Engineering Applications of Artificial Intelligence, 2026. https://arxiv.org/abs/2503.06991
  3. W. Zheng, K. Chen, Y. Guo, Y. Xiao. Classification-Head Bias in Class-Level Machine Unlearning: Diagnosis, Mitigation, and Evaluation. arXiv:2605.08730, 2026. https://arxiv.org/abs/2605.08730

Thank you for providing such a detailed rebuttal. My concern regarding methodological novelty remains. However, other points have been mostly addressed, and I will increase my score to 3.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revised manuscript uploaded, request for response

Dear Reviewers,

Thank you again for your reviews. The discussion period ends soon, so we want to briefly summarize the rebuttals.

We have run all the additional experiments requested and addressed each concern in our individual responses. We have also added every proposed change to the manuscript and uploaded the revised version. All new and changed text is in blue.

The main additions are:

  1. Woken recall on all CIFAR-100 and ViT-S/16 checkpoints (Appendices G and H)
  2. Multi-class forgetting with three forget sets (Appendix I)
  3. Repeated calibration and evaluation splits (Appendix D.4)
  4. Matched references selected on held-out data (Appendix D.6)
  5. Nine-way retain-only references (Appendix D.7)
  6. OOD and noise controls for class specificity (Appendix D.8)
  7. Per-method woken and probe gaps with 95% CIs (Table 8)

If our responses and the revision resolve your concerns, we would be grateful if you would consider updating your score. We are happy to answer any further questions before the discussion period closes.

Sincerely, 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.