acceptodds
Under review as a conference paper at ICLR 2027

The Role of Learning and Memorization in Relabeling-Based Unlearning

Abstract

This work studies how the nature of a response generated by a model impacts the efficiency of relabeling-based unlearning, a common unlearning technique that trains the model to fit an unlearn set (i.e., a dataset that we wish the model to unlearn) with alternative responses to prevent it from generating unwanted outputs that align with the unlearn set. We distinguish between two different ways models can generate undesirable outputs: learning-based generation, where the model learns an underlying rule connecting the input and the response (e.g., social stereotypes), and memorization-based generation, where the model memorizes specific information about a given input (e.g., private information like a phone number). We demonstrate through binary classification and TOFU-LM experiments that relabeling-based unlearning converges more slowly and causes greater temporary retain-accuracy loss for learning-based generation than for memorization-based generation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Privacy, Unlearning, LLM

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

This paper investigates how the distinction between learning-based and memorization-based generation affects the difficulty of machine unlearning. The authors develop a theoretical framework to explain why relabeling-based unlearning may be more challenging when model predictions rely on learned general rules, rather than memorized input-specific associations.

Through controlled binary classification and TOFU-LM experiments, they evaluate several unlearning methods and show that learning-based generation generally requires more optimization steps and causes greater temporary degradation in retain accuracy. Additional experiments examine the effects of different forget-set sizes, model architectures, and held-out paraphrases.

Overall, the paper explores an interesting and meaningful research question and provides useful theoretical and empirical insights into the mechanisms underlying machine unlearning. However, some aspects of the experimental design, particularly the use of replacement labels, warrant further examination to better distinguish the difficulty of forgetting existing knowledge from that of learning contradictory responses.

Critical strengths
  1. Interesting and meaningful research question. The paper investigates an important yet relatively underexplored aspect of machine unlearning: how the underlying generation mechanism (learning-based vs. memorization-based) affects unlearning difficulty. This distinction provides a useful perspective for understanding why certain knowledge may be harder to forget than others. The research question is well motivated and has potential implications for developing more effective unlearning methods.

  2. Combination of theoretical analysis and empirical validation. The paper provides a theoretical framework to explain the different behaviors of learning-based and memorization-based generation during unlearning. The authors further investigate these differences through controlled binary classification and TOFU-LM experiments, considering multiple unlearning methods and additional experimental settings. The combination of theoretical insights and empirical observations strengthens the paper’s contributions and provides a useful foundation for further research.

Critical weaknesses

While the paper presents an interesting and valuable perspective on machine unlearning, I have several concerns regarding the experimental design and the interpretation of the results.

  1. The paper distinguishes learning-based and memorization-based generation using artificially constructed tasks (LINEAR/RANDOM and TOFU-LM). However, real-world LLM knowledge often involves both mechanisms simultaneously, making it unclear whether the findings generalize to realistic scenarios.

  2. The experiments primarily implement unlearning by replacing original answers with incorrect labels. Consequently, the observed difficulty may reflect not only forgetting existing knowledge but also learning contradictory responses, particularly when the original predictions follow a shared rule.

  3. Training models to produce incorrect answers may not be desirable in practical unlearning applications, where maintaining reliability and correctness is important. It would be valuable to investigate whether the observed differences persist when using neutral responses, such as “I don’t know,” instead of incorrect labels.

  4. The evaluation primarily measures whether models stop producing the original answers, which does not necessarily indicate that the underlying knowledge has been removed. This concern is particularly relevant to TOFU-LM, where evaluation is restricted to four candidate answers.

  5. Although the theoretical analysis offers an interesting explanation for the observed differences, the experiments do not directly establish that the slower convergence of learning-based unlearning results from the proposed hypothesis-space expansion. Further empirical evidence would help strengthen the connection between the theoretical analysis and the observed behavior.

Questions for authors
  1. How well do the authors expect the learning-based and memorization-based distinction to generalize to real-world LLM knowledge, where both mechanisms may coexist? Is there any additional evidence supporting the applicability of the findings beyond the artificially constructed tasks?

  2. Would the observed differences in unlearning difficulty persist if incorrect replacement labels were replaced with neutral responses, such as “I don’t know”? This comparison would help clarify whether the results primarily reflect knowledge forgetting or the difficulty of learning contradictory responses.

  3. How do the authors distinguish genuine knowledge removal from response suppression in their experiments? In particular, could the original knowledge remain accessible through alternative prompts or evaluation strategies beyond the four-choice setting used in TOFU-LM?

  4. Is there any empirical evidence directly supporting the proposed hypothesis-space expansion as the explanation for slower convergence in learning-based unlearning? Could alternative factors, such as differences in optimization difficulty or gradient interference, also explain the observed behavior?

2 comments

We thank the reviewer for the positive assessment and the constructive suggestions.

Generalization to real-world knowledge (W1, Q1), real knowledge often involves both mechanisms. Our framework handles this with a mixture over hypotheses (Section 3.1). In a mixture experiment on Qwen3-4B, half of the input space follows a rule and the other half has random labels, and the results fall between the two pure tasks (Appendix C.7.5). Existing unlearning benchmarks do not mark which facts were learned as rules and which were memorized, so they cannot isolate the effect we study. Extending the study to natural data is future work.

Incorrect labels versus neutral responses (W2, W3, Q2), thank you for suggesting this comparison. We ran it on TOFU-LM, replacing the incorrect forget targets with "I don't know" (new Appendix C.6). During scoring we add "I don't know" as a candidate, so a retain question answered with it counts as an error. The gap remains. In the learning-based condition, retain accuracy drops further and recovers later, original answers are suppressed more slowly, and joint convergence takes nearly twice as long as in the memorization-based condition. Two other results agree. NPO, which uses no replacement targets at all, also suppresses original answers more slowly in the learning-based conditions (Table 2). In the new BAND control (W5, Q4), the model fits contradictory labels and still unlearns as fast as on the memorization-based task. Together, these results suggest that the difficulty comes from fitting new responses while a learned rule predicts the old ones.

Suppression versus removal (W4, Q3), we measure unlearning performance primarily with accuracy on the forget and retain sets, a metric widely used in prior LLM unlearning work (e.g., Li et al., 2024; Deeb & Roger, 2024). To test whether suppression generalizes beyond the prompts used during unlearning, we follow Wu et al. (2025) and evaluate held-out paraphrases of each forget question, which were excluded from both fine-tuning and unlearning (Section 4.2). The paraphrases ask for the same attribute in different words, so they show whether the original answer is still reachable from other prompts. In the learning-based condition, held-out accuracy stays elevated after accuracy on the training prompts has approached zero, and held-out unlearning is much slower than in the memorization-based condition (Table 2). This fits our main finding. Complete or certified removal of the underlying knowledge remains an ongoing area and is beyond the scope of this paper. We leave stronger evaluations, such as free generation, to future work.

Evidence for hypothesis-space expansion (W5, Q4), we have added controlled experiment on the binary task (new Appendix C.4). BAND and BAND-MIXED use the LINEAR rule, relabel the same number of points at a similar distance from the decision boundary, and differ only in whether the relabeled points can be fit by moving the boundary. When they can (BAND), unlearning converges about as fast as for memorized labels. When they can only be fit as exceptions (BAND-MIXED), unlearning is slower and retain error roughly doubles. The final decision boundaries on LINEAR show the original rule with exceptions at the forget points (Appendix C.7.1). These results support the view that unlearning slows down when the model has to move from its rule to a rule with exceptions.

Reviewer_qPvROpenReview ↗

Maintain weak accept

The authors' response addresses most of my concerns, I maintain my weak accept recommendation.

Reviewer_ig7SOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks why relabeling-based unlearning is harder for some outputs than others. It separates learning-based generation, where a shared rule produces the output (e.g., stereotypes), from memorization-based generation, where input-specific facts are recalled (e.g., private data). Relabeling is modeled as hypothesis testing with "fabricated evidence" with belief measured by log-posterior odds. Under this model the learning hypothesis loses belief linearly in evidence length, while the memorization hypothesis’s loss is bounded. Experiments on binary classification and a modified TOFU show that memorization-based generation unlearns faster and with less retain-accuracy loss.

Critical strengths
  1. The learning vs. memorization distinction is intuitive and useful. It differs from per-example memorization scores [1].

  2. The “fabricated evidence” hypothesis-testing framing is clean and easy to follow.

  3. The experiments are well controlled: the two conditions differ only in how labels are assigned. The gap holds across two relabeling methods, three models, several forget-set sizes and held-out paraphrases.

  4. The paper is clear about its scope. It reports where the pattern breaks, such as the NPO reversal on Qwen3-4B.

[1] Zhao, K., et al. What makes unlearning hard and what to do about it. NeurIPS 2024.

Critical weaknesses
  1. Clarity: At first it is not obvious how the two formalizations differ. The real difference is in the prior: H1 puts an i.i.d. uniform prior on the entries of V, while H0 is a point mass on f.

  2. Connection to exposure. The test statistic is the log-posterior odds. It would be nice to connect it to exposure [2], and to the generalized exposure used for unlearning evaluation [3]. Both are likelihood-ratio or rank-based measures of how strongly a model favors the original response over alternatives.

  3. Unclear why stability, meaning a bounded log-odds change relative to constant baselines, is the right quantity, and how it connects to convergence speed and retain accuracy under gradient-based optimization. Right now the link from Theorems 2 and 3 to the empirical metrics is only informal.

  4. Theorem 4: scope. The result is a worst-case existence result about the geometry of halfspaces, specifically a relabeled point inside a same-label cluster. It does not address convergence speed. Please temper the claim in Section 3.2 that it “helps explain the slower convergence”.

  5. The evidence is limited in scope. Both tasks are synthetic: 2D binary classification with 1k points, and a single TOFU attribute with 4 labels scored by constrained likelihood. Models are 3–8B. Results are means over 3 runs with no variance reported. A single learning rate and retain weight are used, even though convergence-speed comparisons are sensitive to both.

  6. The baselines H2/H3 are probably already ruled out by the history h. A larger drop for H0 relative to them doesn’t show whether the model adopts a hypothesis that fits both U' and R.

Theorem 4 proof:

  • "Negligible" is undefined, and P(x = x′) = 0.2/3^d is only small when d is large.
  • What is unit B?

[2] Carlini, N., Liu, C., Erlingsson, Ú., Kos, J., Song, D. The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networks. USENIX Security, 2019.

[3] Baluta, T., Lamblin, P., Tarlow, D., Pedregosa, F., Dziugaite, G. K. Unlearning in- vs. out-of-distribution data in LLMs under gradient-based methods. NeurIPS 2024 Workshops (SafeGenAI). arXiv:2411.04388.

[4] Allouah, Y., Kazdan, J., Guerraoui, R., Koyejo, S. The Utility and Complexity of In- and Out-of-Distribution Machine Unlearning. ICLR 2025. arXiv:2412.09119.

Questions for authors
  1. How do Theorems 2 and 3 translate into the number of gradient steps to convergence and the retain-accuracy drop? Is there a formal or empirical link beyond the qualitative argument?

  2. Why compare against the constant baselines H2/H3? Would the separation still hold against a hypothesis that fits both U′ and R, for example f with an exception at x*?

  3. Given 20 epochs of fine-tuning, how large is |m₁ − m₋₁| for forget examples in your experiments, and how does Theorem 3’s bound compare to Theorem 2’s drop at the evidence lengths you actually use?

  4. How does the learning/memorization distinction relate to in- vs. out-of-distribution forget examples [3, 4]? Is a learning-based example essentially an in-distribution one, as the construction in Theorem 4 suggests?

  5. Can the log-posterior odds be related to exposure [2, 3], and could exposure be reported as an additional metric in the experiments?

3 comments

We thank the reviewer for the careful reading and the precise questions.

Difference between the two hypotheses (W1), the reviewer is right that the difference lies in the prior. H0H_0 places all prior mass on ff, while H1H_1 places an i.i.d. uniform prior over VV. We have revised §3.1 to state this when these two hypotheses are first introduced. Under H0H_0, the label at one input determines the labels at other inputs through ff. Under H1H_1, the labels at different inputs are independent.

Connection to exposure (W2, Q5), exposure [2, 3] compares the likelihood that one model assigns to the original response with the likelihoods of alternative responses, whereas our log posterior odds compare two hypotheses about how the responses were generated. These two are linked through the predictive distribution in §3.1. In our setting, the analogue of exposure is the model's log-odds of the original response over the relabeled response. We have added this connection to the paper and leave reporting exposure as a metric to future work.

Connection between the theory and convergence (W3, Q1), our theory makes two testable predictions. First, under H0H_0, relabeled evidence counts against the rule that also governs the retain labels, so relabeling should disturb retain predictions across the input space, while under H1H_1, it affects one input only. Second, fitting both the relabeled and the retain data also requires moving to a hypothesis outside the original rule class, such as ff with exceptions (Theorem 4). Figures 6, 8 and 10 confirm both effects across all three models. As mentioned by the reviewer, the theorems do not directly bound the number of gradient steps, and we have clarified this point in the revision.

To link hypothesis expansion to convergence speed directly, we added a controlled experiment (new Appendix C.4), comparing BAND and BAND-MIXED. Both relabel the same number of LINEAR points at a similar distance from the decision boundary, providing the same evidence against the original rule. New labels in BAND can be fit by a shifted linear rule, while BAND-MIXED requires a rule with exceptions. Empirically, BAND converges nearly as fast as RANDOM, while BAND-MIXED converges more slowly, and with roughly twice the retain error. This suggests that the slowdown comes from moving to a hypothesis outside the original rule class.

Scope and proof of Theorem 4 (W4, W7), Theorem 4 is a worst-case existence result and does not address convergence speed. We have tempered the corresponding claims in §3.2 and the introduction. In the proof, we corrected BB to CC (the cube) under the ℓ∞\ell_\infty norm. The probability P(x=x′)=0.2⋅3−dP(x = x') = 0.2 \cdot 3^{-d} decays exponentially in dd, and dd is large in most modern high-dimentional ML settings, so it is negligible in practice. To cover small dd as well, we now state this probability explicitly in the theorem, together with the bound errD(h^)≥0.1(1−3−d)\mathrm{err}_D(\hat{h}) \geq 0.1(1 - 3^{-d}) for every dd.

Scope and strength of the evidence (W5), we have added standard deviations for Table 2 in a new Table 3. For both relabeling methods and both tasks, the differences between conditions in retain error and convergence time are larger than one standard deviation of either condition. We have also added a sweep over learning rates and retain weights for Standard Relabeling on the binary task (new Appendix C.8). In every setting where unlearning converges within the budget, LINEAR takes longer than RANDOM. Regarding benchmark choice, we constructed synthetic tasks on purpose because existing unlearning benchmarks do not distinguish between rule-learned and memorized facts. And natural data usually entangles both, preventing us from isolating the mechsnism we study. Extending the analysis to larger models and natural data is a direction for future work.

Constant baselines (W6, Q2), Theorems 2 and 3 depend on the reference hypothesis only through its prediction at x∗x*, because Δe\Delta_e equals the log-likelihood ratio of ee (Eq. 2), and ee consists of observations at x∗x* only. Consequently, any hypothesis that predicts the relabeled label at x∗x*, including ff with an exception at x∗x*, yields the same Δe\Delta_e, so the separation holds against it. We agree with the reviewer that the constant baselines are ruled out by this history hh, so it is informative that we claim they can dominate H0H_0 in the Remark. We have rewritten the Remark using ff with exceptions as the reference hypothesis, which fits both U′U' and RR. Its posterior overtakes H0H_0 once the relabeled evidence exceeds the label imbalance at x∗x* plus a prior-odds term, aligning with the final decision boundaries observed in Figures 6–10.

In- and out-of-distribution forget data (Q4), Both learning-based and memorization-based forget examples are in-distribution: the former follows directly from the construction of Theorem 4, and the latter holds because forget and retain points are generated by the same labeling process within each condition. The learning/memorization distinction is therefore separate from the in-/out-of-distribution distinction studied in [3, 4]. It characterized whether labels are generated by a shared rule or independently for each input. We have added both references to our related work section.

I would like to thank the authors for their rebuttal. It addressed my questions and I believe the remaining concerns can be addressed through the revised manuscript. I will raise my rating from 2 to 3.

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

The paper studies how the efficacy of relabeling-based unlearning can depend on whether the model has learned the unwanted data through rule-based learning (e.g., social stereotypes) or specific memorization (e.g., private information such as phone numbers). Regarding relabeled data points as fabricated evidence, the paper leverages a Bayesian hypothesis testing framework to theoretically bound the belief change, which quantifies the belief of a hypothesis (learning-based vs. memorization-based) to a baseline hypothesis that makes constant predictions. The study shows the learning-based hypothesis is less stable under conflicting evidence. The experiments test the LLaMA-3.2-3B-Instruct model on (1) a synthetic binary classification task and (2) an extended version of the TOFU benchmark to show that learning-based conditions lead to larger retain-knowledge degradation.

Critical strengths
  • [S1] Interesting perspective on unlearning. Studying how the way the LLM reproduces knowledge (through rule-based reasoning or verbatim memorization) affects unlearning efficacy is an interesting take, and would be of interest to the unlearning research community. As many existing works study different unlearning targets such as different domains (PII or harmful information) or knowledge granularity (entity-level or fact-level), considering the effect of the model's own generative process adds another useful axis. The extended version of TOFU is also a nice addition providing a good testbed for this concept.
  • [S2] Consistency in experimental results. The main empirical findings appear consistently across (1) both binary classification and TOFU tasks, (2) a total of three model families (LLaMA-3.2-3B-Instruction, and LLaMA-8B/Qwen3-4B shown in the appendix), and (3) different forget set sizes. That being said, the experimental scope is fairly broad and demonstrates the generalizability of the results.
Critical weaknesses
  • [W1] Narrow scope. The analysis in the paper is strictly limited to relabeling-based unlearning scenarios, and because it is yet unclear whether relabeling achieves meaningful unlearning in the first place, this comes as a main weakness. Previous work showed that knowledge removed by tuning-based unlearning methods can be recovered with simple attacks such as rephrasing or quantization [A, B]. This is also shown in the Held-out evaluation section of the paper. In addition, overwriting unwanted knowledge with fabricated evidence deviates from what a gold-standard retrain-from-scratch model would show, which is to spread probability across multiple plausible responses [C]. Analyzing not only relabeling-based methods, but also other unlearning methods through the lens of rule learning vs. memorization would greatly strengthen the paper, though that would require a major update.
  • [W2] Unclear relevance of theoretical results to unlearning. Despite the interesting theoretical studies, the connection between the theoretical results to actual unlearning effectiveness is quite blurry. For instance, the log 2 upper-bound in Theorem 3 is valid when x^* does not appear in h, but if x^* is contained in the forget set, that inevitably means that x^* did appear in h, so the log 2 bound never happens. Furthermore, Theorem 4 discusses the limited expressivity of linear models, claiming that they cannot isolate out the relabeled sample from others. The models with which unlearning is relevant are much larger and much more expressive, so it is unclear how Theorem 4 can be useful in explaining unlearning behavior on such models.
  • [W3] Potential confounder in the LINEAR vs. RANDOM results. In the LINEAR data setup for binary classification and L condition for TOFU-LM, the relabeled forget data points directly contradict the nearby retain data points. Therefore by construction, trying to relabel breaks the prediction on nearby retain data. This is not shown in the RANDOM setup and M condition, because the labels are independent, and this entanglement between the forget data and nearby retain data does not exist. In all, this implies that the gap in unlearning efficacy could be due to how much the relabeled forget samples are entangled to nearby retain samples, rather than whether the model knows the forget samples through learning or memorization. Unless this entanglement aspect is isolated out in the experiments, the experimental findings seem somewhat thin.

[A] Lynch et al., Eight Methods to Evaluate Robust Unlearning in LLMs. arXiv 2024.
[B] Zhang et al., Catastrophic Failure of LLM Unlearning via Quantization. ICLR 2025.
[C] Cho et al., Reference-Specific Unlearning Metrics Can Hide the Truth: A Reality Check. arXiv 2025.

Questions for authors

Any clarifications on the points in the weaknesses block above would be helpful in understanding the draft better.

2 comments

We thank the reviewer for the thoughtful comments.

Scope of relabeling-based unlearning (W1), we chose to study relabeling because it is widely used in practice (e.g., Maini et al., 2024; Eldan & Russinovich, 2023; Mekala et al., 2025). Therefore it is useful to understand when it converges slowly or loses retain accuracy. Relabeling is also tractable for formal analysis and for controlled experiments. Furthermore, we evaluate two non-relabeling methods, NPO and RMU (Figures 3(c–d) and 5, Table 2), and both show a similar gap to some degree. Extending our theoretical framework to these methods is a promising direction for future work.

Regarding robustness, the reviewer's concerns apply to fine-tuning-based unlearning in general. For example, quantization can restore knowledge removed by gradient-ascent and NPO variants [B], and many methods with near-optimal standard scores still diverge from a model retrained without the forget dataset [C]. While these are important problems, a single paper cannot address all of them. In this work, we focus on one dimension: unlearning efficiency, and study it within a formal framework.

Relevance of the theoretical results (W2), we use a simplified setting to keep the analysis tractable. Theorems 2 and 3 show that relabeled evidence keeps lowering belief in a learned rule, while belief in memorized labels drops by a bounded amount. Together with Theorem 4, they predict that relabeling changes predictions globally across the entrie input space for a learned rule, but only locally for memorized labels, as Figures 6, 8 and 10 show.

We appreciate the reviewer noting that the log 2 bound does not apply to forget points. Instead, they are governed by the general upper bound in Theorem 3, which is determined by each point's label counts in the training history. This bound remains fixed with the increasing relabeled evidence, whereas it grows linearly under the learning hypothesis.

Large models eventually fit relabeled points as exceptions (Figures 6(c), 8(d) and 10(d)). Theorem 4 illustrates why a learned rule must be revised to fit such points. A new controlled experiment (new Appendix C.4) shows that this revision slows unlearning in LLMs, since relabeled points that require exceptions are unlearned more slowly than points a shifted rule can fit. We have revised the manuscript to clarify that Theorem 4 serves as a conceptual illustration of this mechanism.

Entanglement between forget and retain data (W3), thank you for pointing out this potential confound. The original comparison does not decouple the presence of an underlying rule from local label conflicts with nearby retain points, so we added a controlled experiment (new Appendix C.4). In RANDOM+, each forget point shares its original label with its nearest retain neighbors, so relabeling induces local conflict with nearby retain data just as in LINEAR, but without any learned rule. RANDOM+ converges nearly as fast as RANDOM, while LINEAR stays much slower. This shows local label conflict accounts for very little of the observed gap. Since a learned rule couples predictions across the entire input space, relabeling under LINEAR also affects retain points far from the forget points.

AuthorsOpenReview ↗

Follow-up on our response

Thank you again for your thoughtful review. We hope you have had an opportunity to consider our detailed response and the revised manuscript. These include the added RANDOM+ comparison in Appendix C.4, which specifically addresses your concern about local conflicts between forget and retain examples, along with clarifications of the theoretical results and the scope of our analysis.

As the discussion period draws to a close, please let us know if any questions or concerns remain that we could help clarify. We would be happy to provide further clarification. We hope these additions and explanations have addressed your concerns. If they have resolved the concerns underlying your initial assessment, we would greatly appreciate your reconsidering and updating your review and rating accordingly. Thank you again for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank all reviewers for their careful and constructive reviews. The reviewers found the learning versus memorization question interesting and useful (Reviewers URvp, ig7S and qPvR), the fabricated-evidence framing clean (Reviewer ig7S), and the experiments well controlled and consistent (Reviewers URvp and ig7S). Below we summarize how we addressed the reviewers' main concerns. All changes are marked in blue in the revised PDF.

Entanglement between forget and retain data (Reviewer URvp). To separate the shared rule from local conflict with nearby retain data, we added RANDOM+, in which each forget point's nearest retain neighbours share its original label, so relabeling creates local conflict without a rule. RANDOM+ converges about as fast as RANDOM, while LINEAR remains much slower (Appendix C.4).

Connection between hypothesis expansion and convergence speed (Reviewers ig7S and qPvR). BAND and BAND-MIXED keep the LINEAR rule and vary only whether the relabeled points can be fit by moving the decision boundary or must be fit as exceptions. Unlearning is about as fast as on RANDOM in the first case and clearly slower in the second, which links hypothesis expansion to convergence speed (Appendix C.4).

Hyperparameter sweep and standard deviations (Reviewer ig7S). We added standard deviations for the main results (Table 3) and a sweep over learning rates and retain weights (Appendix C.8). In every setting where unlearning converges, LINEAR takes longer than RANDOM.

Relabeling with neutral responses (Reviewer qPvR). Replacing the incorrect targets with "I don't know" on TOFU-LM leaves the gap in place. The learning-based condition still loses more retain accuracy and converges more slowly (Appendix C.6).

Clarified theory and proofs (Reviewers URvp and ig7S). Section 3.1 now states that H0H_0 and H1H_1 differ in their prior, explains why the choice of reference hypothesis does not affect Theorems 2 and 3, and connects our log posterior odds to exposure. The Remark after Theorem 3 now uses ff with exceptions as the competing hypothesis, notes that the general bound in Theorem 3 applies to forget points, and states that the theorems are qualitative and do not give a number of gradient steps. We present Theorem 4 as an illustration of why a learned rule must be revised, soften the related claims in Sections 1 and 3.2, and clarify its proof in Appendix B.

Scope and positioning (Reviewers URvp, ig7S and qPvR). Our response to Reviewer URvp explains the focus on relabeling and its relation to the non-relabeling methods we also evaluate. The related work in Appendix A now discusses in- and out-of-distribution unlearning and how our distinction complements it. The conclusion lists evaluation beyond constrained label scoring, such as free generation, as future work.

We hope these additions address the reviewers' concerns. Please let us know if any further comments or questions come up, and we thank the reviewers and the Area Chair again for their time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.