acceptodds
Under review as a conference paper at ICLR 2027

Conformal Unlearning: A New Paradigm for Output Control Under Conformal Prediction

Abstract

Existing machine unlearning definitions generally constrain the objective of unlearning to approximating a model retrained from scratch after removing the forget set. However, this formulation equates unlearning to `untraining' under all scenarios, presenting substantial practical limitations as it often neglects the gap between parameter-space proximity and output-level consistency, is vulnerable to falsification, and overstates retrained models as a universal gold standard for unlearning. To address these issues, we introduce conformal machine unlearning, a new paradigm that does not require a retrained model as the reference, and instead focuses on controlling the output of the model. We formalize conformal unlearning as an output-control framework that enforces high conformal coverage on retained data and high conformal miscoverage on forgotten data, develop empirical evaluation metrics, and present an algorithm that optimizes these conformal objectives. Extensive experiments on vision and text benchmarks verify that the proposed method effectively removes targeted information while preserving utility.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Machine unlearning, conformal prediction, fake unlearning, output control

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

The paper proposes conformal machine unlearning, a framework that defines successful unlearning through the behavior of conformal prediction sets rather than similarity to a model retrained from scratch. The authors introduce EFACE (Excluding Forgotten data by Adjusting CoveragE), an optimization-based method designed to increase miscoverage on the forget set while preserving coverage on retained samples. The method is evaluated on ImageNet100, CIFAR100, and 20NewsGroups, considering both class-wise and cluster-wise forgetting scenarios.

Critical strengths

The paper's motivation that a drop in top-1 accuracy may not be sufficient evidence of forgetting, and that a model can predict the wrong label while still including the true label in a conformal prediction set, seems interesting. The paper derives a connection between α\alpha and β\beta in the (α,β)(\alpha,\beta)-conformal unlearning algorithm through Lemma 2 under the mentioned assumption. The evaluation includes several established unlearning baselines, in-sample and out-of-sample unlearning, class-wise and cluster-wise forgetting, and multiple datasets.

Critical weaknesses

(1) Due to the strong assumption noted in Line 149, the proposed algorithm cannot handle random forgetting scenarios and relies heavily on access to the forget set. (2) The cluster-wise forgetting setting lacks sufficient justification as to why it serves as a practical primary evaluation scenario. (3) Although the paper introduces novel unlearning metrics, the authors must still demonstrate how conformal machine unlearning performs under standard metrics, including forget accuracy (FA), retain accuracy (RA), membership inference attack (MIA), and test accuracy (TA). The limited MIA results presented in Table 7 fail to establish the superiority of the proposed approach. Furthermore, the newly defined metrics in the paper rely on maximum prediction-set sizes (the thresholds c for retained data and d for forgotten data). Thus, results obtained using these new metrics may appear strong without demonstrating that the desired behavior holds across most samples. (4) The authors should clarify the sensitivity of their theoretical guarantees and empirical results when distribution shift occurs between the calibration set and the forget cohort. Additionally, the authors should explicitly quantify the coverage gap across all evaluated settings. (5) Deferring the entire Related Work section to the appendix is bad practice and makes it difficult to assess the paper's novelty relative to existing work.

Questions for authors

Please see the weaknesses.

4 comments
AuthorsOpenReview ↗

Follow-up on our responses

We thank the reviewer again for the thoughtful feedback and questions. We appreciate that the reviewer found our motivation interesting: a drop in top-1 accuracy may not be sufficient evidence of forgetting, because the true label can remain in the conformal prediction set. We also appreciate the reviewer's recognition of our evaluation across several baselines, datasets, and forgetting scenarios.

The reviewer noted only "minor or fixable concerns about soundness", and we addressed each of them in our response and the revision:

  • We clarified the distinguishability assumption and the use of out-sample proxy data.
  • We explained why we evaluate on clusters.
  • We added accuracy and MIA results for all experiments.
  • We referred to our analysis under distribution shift (Appendix E).
  • We moved the related work to the main text (Section 6).

These changes improved the paper and its positioning. We would appreciate it if the reviewer could confirm that our responses resolve these concerns, or tell us what remains and we will be happy to address it.

Reviewer_dAFYOpenReview ↗

Official Comment by Reviewer dAFY

I thank the authors for the provided responses. Some of my concerns still remain.

  • I understand the motivation behind unlearning and untraining, and I believe this is an interesting area of research. However, the paper is written in a way that makes its motivation and novelty relative to existing work unclear. The Related Work section (Section 6) does not read like a proper literature review. Additionally, several recent works, such as [1], have not been discussed in the paper.

[1] E. Triantafillou, A. I. Humayun, M. Ribero, A. M. Turner, M. C. Mozer, G. Kaissis, "Is your algorithm unlearning or untraining?" arXiv preprint arXiv:2604.07962, Apr. 2026.

  • I thank the authors for adding the accuracy and MIA results to some of the tables. However, unlike the metrics you have defined, accuracy and MIA are meaningful evaluation metrics when their results are compared against those of the gold-standard method, namely, retraining from scratch. However, the corresponding values for the RT method are missing from the tables.

  • My question regarding the newly defined metrics in the paper, which rely on maximum prediction-set sizes (the thresholds (c) for retained data and (d) for forgotten data), remains unanswered. In particular, the results obtained using these metrics may appear strong without demonstrating that the desired behavior holds across most samples.

AuthorsOpenReview ↗

Follow-up discussion

We thank the reviewer for their engagement with our rebuttal and their further constructive comments.

1. The related work section (Section 6) and discussing Triantafillou et al. (2026).

We respectfully note that the related work section (now titled Comparison with Closely Related Methods to avoid confusion) is dedicated to the most relevant work of Shi, Liu, and Wang (2025), which discusses conformal objectives in unlearning. We contrast our work with theirs and present a theoretical, algorithmic, and empirical critique of theirs. We further refer to residual knowledge effects and to the use of external data to induce unlearning, which bear similarities to EFACE. The literature survey, on the other hand, is part of our introduction where we categorise unlearning methods and explain that they target the untraining objective. For that matter, the naming 'untraining' vs 'unlearning' is due to Triantafillou et al. (2026), which we cite twice in our introduction. Following the reviewer's comment, we now contrast our work with Triantafillou et al. (2026) in the introduction (highlighted in blue, lines 93-98): they also separate unlearning from untraining, but they sketch both notions relative to a retrained model and leave the formal definition and the evaluation of unlearning open, whereas we define the unlearning of a population at the output of the model, with no retrained reference. We are happy to discuss and include any particular relevant work that the reviewer believes would improve the positioning of our work.

2. RT in accuracy and MIA.

Kindly note that we already include the accuracy of RT as well as the certified method PABI and now also report MIA corresponding to all image experiments in Tables 3 (main text), 15, 16, 17, and 18 (all in Appendix H.2.6), which separate our comparison with RT from the comparison with the other baselines. We comment on these results in Section 5.1.2. Nonetheless, we agree with the reviewer that it is useful to consolidate the results with RT along with the other baselines, and we now add those results to the tables of accuracy and MIA. We appreciate the opportunity to clarify that, as explained in our introduction, the case of unlearning populations of data with shared characteristics requires expanding the evaluation approaches, and hence the aim of EFACE is not to get as close as possible to RT in terms of accuracy. This argument is at the core of the discussion our work presents.

3. Reliance on maximum prediction-set sizes in the proposed metrics

This is an important point, and we would like to clarify it. The reviewer is right that CF and mCF are conditional on the set size, so they could appear strong if many samples had sets larger than c or d. Our work addresses this in three ways. First, at c = d = ∣Y∣|\mathcal{Y}|, CF and mCF reduce to the standard coverage and miscoverage over all samples (Section 4.1). Figures 5, 6, and 7 (Appendix H.2.4) report HH for all methods as c and d vary up to ∣Y∣|\mathcal{Y}|, where EFACE remains stable. Second, Corollary 1 (Section 4.1) makes this dependence explicit: the bounds on CF and mCF degrade exactly by the probability that the set size exceeds c or d, so the guarantees account for the excluded samples. Third, smaller values of c and d expose failure modes rather than hide them, such as the collapse of ∇τ\nabla\tau at c = d = 50 in Table 2 (lines 414-417), which is why we recommend evaluating at several levels of c and d (Appendix J.3). Figure 13 (Appendix J) further shows that the average set sizes of EFACE stay small on the retained data and grow mainly on the forget data, while remaining below the evaluated level c = d = 50.

We thank the reviewer again for their time and effort, and we believe the changes stemming from this discussion meaningfully improve our submission. We hope these comments address the raised questions.

Reviewer_dAFYOpenReview ↗

Official Comment by Reviewer dAFY

I thank the authors for their responses and the newly added modifications. I have increased my score to 3 (Weak Accept).

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

Traditional machine unlearning (MU) methods generally rely on approximating a model trained from scratch without the forget data, denoted as RT. However, this RT model might not always be a gold standard, since it can be prone to falsification. To mitigate this, the authors propose conformal machine unlearning. This approach focuses on the output of the model, and enforces high conformal coverage on retain data and high miscoverage on the forget set. Experiments on vision and text tasks are conducted to show the effectiveness of the proposed method.

Critical strengths
  • The paper proposes an approach for conformal machine unlearning which can preserve retain data coverage while achieving strong forgetting.
Critical weaknesses
  • In its current state, the main text is not complete on its own, and it is unclear in the main text how EFACE is different from similar recent literature such as CPU, and related work is only discussed in the Appendix. Related to this, it is not clear what Retraining (RT) means in Table 1. If we assume it refers to a model directly retrained on data excluding the forget set, it should be the gold standard since such model has never seen the forget data. It would help if the authors clarify what RT refers to, so that the motivation can be more clear. Additionally, the main algorithm, which is an approximation of the true objective, is not discussed in the main paper. While these discussions are included in the appendix, the main text is hard to follow, as the motivation, novelty, and algorithm details are not fully discussed. I would suggest rewriting the paper to include important details from the appendix in the main text.

  • The main results in Table 2 only report the metrics defined in the paper. It would be helpful to also include existing MU metrics (e.g. UA, RA, TA, MIA), to provide a comparison. Additionally, CR (Shi et. al, 2025), is not included in Table 2, which is a relevant and recent baseline. While a comparison with CR is discussed in the Appendix, including it in the main paper helps clarify the advantages of the proposed method. A discussion can also be included to see where the traditional metrics disagree with the proposed ones.

Questions for authors

Can the authors provide an explanation on what Retraining refers to, where the method stands compared to recent literature, and provide extended comparison with other metrics?

3 comments
AuthorsOpenReview ↗

Follow-up on our responses

We thank the reviewer again for their feedback. We appreciate the reviewer's recognition that our approach preserves coverage on the retained data while achieving strong forgetting, and that the reviewer has "no substantive concerns about the paper's soundness".

The remaining concerns were about the completeness and readability of the main text, and we addressed them in the revision:

  • The related work and the comparison with CPU and CR are now in the main text (Section 6).
  • The practical implementation of EFACE is now in the main text.
  • We clarified that RT follows its standard definition.
  • We added accuracy and MIA results for all coverage-miscoverage tables.

The reviewer's suggestions made the paper clearer and better positioned. We would appreciate it if the reviewer could acknowledge our responses and consider whether they change the evaluation, or tell us what remains unclear. This discussion is the best time to address it, and we are happy to answer any remaining questions.

Dear Reviewer AcEk,

We are grateful for your thoughtful comments and for the time you have dedicated to reviewing our paper.

As the end of the discussion phase is near, we would like to briefly follow up on our responses and the revisions we made. We hope they address the questions you raised.

If you have had a chance to go through our replies, we would very much appreciate knowing whether anything remains unclear or unresolved. We would be happy to provide any further details while the discussion is still open.

Many thanks for your consideration.

Kind regards,

The Authors of Paper 9131

I thank the authors for their response. I believe the revised version is improved compared to the original. Additionally, I believe it would be helpful to include some class-wise results in the main text if possible, since they are only presented in the appendix. Regarding the final recommendation, I will consider the revisions and any discussions with the other reviewers and AC.

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

The work posits that comparison to a retrained model is limited in that there may exist parameter space distributions that are distinct between the unlearned model and the RT baseline and vice versa, in addition to a problem of scalability and availability of the RT baseline, and the phenomenon of "fake conformal unlearning" wherein the RT baseline retains substantial distributional information about the unlearned set despite metrics based performance over it. It makes a conformal unlearning proposition wherein unlearning is guided by coverage and mis-coverage probabilities that is independent of a RT baseline and lead to uncertainty quantification, motivated by theoretical work that formulates the proposed unlearning objectives and assumptions, culminating in a probabilistic view of unlearning and a ERM based objective and evaluation metrics. Specifically, it proposes (α,β)(\alpha, \beta)-conformal unlearning wherein after unlearning, conformal prediction sets must cover retained data with probability at least 1−α\alpha and miss the true label on forgotten data with probability at least β\beta. The forget set is defined by a shared characteristic (a class, a cluster, a feature), not individual samples, and the paper assumes the forget and retain distributions are distinguishable.

Critical strengths
  1. The view on evaluation of unlearning is valid and demonstrated. A model could show a low top-1 accuracy on the forget data while its prediction sets still contain the true label. Set-valued coverage makes for a useful evaluation setup to ensure comprehensive evaluation especially when the current set of metrics disagree, to allow the reader for a comprehensive view of the capacities and limitations of the methods and the network.

  2. The metrics are simple and the harmonic mean over six subsets is a fair way to penalize methods that only forget the exact samples they were given.

  3. The experimental coverage is broad: ten baselines, three datasets, in-sample and out-sample unlearning, class-wise and cluster-wise, plus retraining and a certified method, with re-conformalization applied to everyone so the comparison is fair on the conformal side.

  4. The theoretical coverage when interpreted helps with a deeper understanding of the proposition and unlearning in general.

Critical weaknesses
  1. The central premise that the retrained model isn't the gold standard isn't supported, and an alternative one could be the criticism of the widely used evaluation metrics which could help motivate the contribution better. The issues raised (parameter closeness doesn't imply behavior closeness and parameter closeness can be forged) are fair criticisms of the evaluation measures, not of RT itself. Having the forgotten samples' true labels within the conformal prediction set may be an artifact of semantic similarity or generalization which may be desirable. The work could be reframed as conformal metrics being a useful addition to comparing against other methods and the retrained model, not a replacement for it. This is supported by the widely made GDPR based motivation that calls for sample specific erasure which this work cannot support any more than existing methods which also report population-level metrics, while the retrained model does give a per-sample guarantee in that the sample was never used.

  2. The proposed method EFACE is gradient ascent on the forget set plus gradient descent on the retain set with a hinge at the running α-quantile (Eq. 24a which should be a part of the main body for clarification). The quantile threshold is the only new piece that may be missing ablations, limiting the assessment of the value of the incremental term.

  3. Miscoverage as defined in the work is constrained by the size of the forget set. The authors could clarify the attainable miscoverage in tables 2, 3 and 4. Calibrating on the retain data fixes this? If so, it could be a more natural choice for the experiments? The authors could clarify.

  4. Sample specific forgetting is a strong use-case of the field that the work cannot cover as it is formulated currently. When the forgotten samples come from the same distribution as the retain set, mis-coverage on them cannot exceed α\alpha without breaking coverage on the retain set, so the definition is only satisfiable for group-level forgetting. An adaptation could require that the coverage on the Df\mathcal{D}_f to match on the held out data (at 1−α1-\alpha), akin to the retrained model.

Questions for authors
  1. What were the top-1 prediction rates or empty set rates for the forget set for the methods? This contextualizes the contribution since the right prediction could fall out of the threshold despite being top-1, thereby making a case for the existing metrics.

  2. Why not calibrate on the retain set only? Doesn't it relieve the constraint of β≤α/πf\beta \le \alpha/\pi_f?

  3. What does the sigmoid hinge add over the gradient ascent over forget set and gradient descent over the retain set objectives that several existing works use? An ablation would be helpful.

  4. Is the backbone model different for the baseline methods in Tables 2, 3? Could ResNet 18 be incorporated into the baselines with the same c, d settings so as to remove this confounding factor in performance assessments?

  5. Several baselines in Table 2 already fail on accuracy alone, so the conformal metric is most useful where it disagrees with accuracy. It would help to highlight those cases.

2 comments
AuthorsOpenReview ↗

Follow-up on our responses

We thank the reviewer again for the thoughtful and constructive feedback. We appreciate that the reviewer found our view on unlearning evaluation valid and demonstrated, in particular that a model can show low top-1 accuracy on the forget data while its prediction sets still contain the true label. We also appreciate that the reviewer found our experimental coverage broad and fair.

The reviewer noted only "minor or fixable concerns about soundness", and we addressed each of them in the revision:

  • We reframed our positioning: conformal metrics complement retrain-based evaluation and do not replace it.
  • We added an ablation against gradient ascent-descent (Appendix I).
  • We stated the attainable miscoverage in all table captions.
  • We added accuracy tables and used the same backbone across all comparisons.

These changes meaningfully improved the paper, and we are grateful for them. We would appreciate it if the reviewer could confirm that our responses resolve these concerns, or tell us what remains so we can address it.

Dear Reviewer eXKt,

Thank you again for your careful review and the constructive suggestions on our submission. We appreciate the time you have given to this process.

With the discussion period coming to a close, we wanted to check in on our earlier responses and the updates we made in the paper. We hope these have resolved the points you raised.

If you have had the opportunity to look at our responses, we would be grateful to hear whether any of your concerns still stand. We are glad to clarify anything further before the discussion ends.

Thank you for your time and consideration.

Best regards,

The Authors of Paper 9131

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Key Updates

Dear Area Chair,

Thank you for the time and effort you have put into overseeing the review of our submission. We are grateful to all reviewers for their thoughtful and constructive feedback, and for recognising the main strengths of our work: a valid and demonstrated view on the evaluation of unlearning, where low top-1 accuracy can hide true labels that remain in the prediction sets (eXKt, dAFY); a method that preserves coverage on the retained data while achieving strong forgetting (AcEk); and a broad evaluation with ten baselines, three datasets, and in-sample, out-sample, class-wise, and cluster-wise settings (eXKt, dAFY). None of the reviewers raised a major soundness concern: Reviewer AcEk reported no substantive concerns, and Reviewers eXKt and dAFY reported only minor or fixable ones.

During the rebuttal period, we addressed all raised concerns and updated the manuscript accordingly (changes highlighted in blue). We are confident that our responses and revisions answer every question the reviewers raised, and they have made the paper meaningfully better. Our main updates are:

  • Positioning (eXKt, AcEk, dAFY): Conformal unlearning is now framed as a complement to retrain-based evaluation for populations of data. The introduction includes the literature survey and a contrast with Triantafillou et al. (2026), and the comparison with CPU and the CR metric is now in the main text (Section 6).
  • Completeness of the main text (AcEk): The practical implementation of EFACE is now in the main text, and we clarified that RT follows its standard definition.
  • Ablation (eXKt): A new comparison with gradient ascent-descent (Appendix I) shows that, at the same retain accuracy, EFACE reaches much higher miscoverage on the forget set (e.g., 0.92 vs. 0.31), whereas gradient ascent diverges until its prediction sets become trivial.
  • Standard metrics (AcEk, dAFY): Accuracy and MIA results now accompany every coverage-miscoverage setting and include RT (Tables 6–10). The comparison with RT and PABI also reports MIA (Tables 3 and 15–18).
  • Consistent backbone (eXKt): All comparisons now use RepVGG-A2 (Tables 1, 3, 4, and 15–18), and the conclusions are consistent.
  • Attainable miscoverage and set sizes (eXKt, dAFY): The table captions report the attainable miscoverage β\beta, and we explain why calibrating on the retained data alone would remove the guarantee on the forget data. We also clarify that CF and mCF reduce to standard coverage and miscoverage at c=d=∣Y∣c = d = |\mathcal{Y}|, that Corollary 1 bounds the effect of excluding large sets, and that EFACE remains stable across cc and dd (Appendix H.2.4).

We believe these additions have strengthened the paper and resolved the reviewers' concerns, and we hope you will consider them favourably in your final recommendation.

Sincerely,

The Authors of Paper 9131

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.