acceptodds
Under review as a conference paper at ICLR 2027

Robust LLM Unlearning Against Relearning Attacks: The Minor Components in Representations Matter

Abstract

Large language model (LLM) unlearning aims to remove specific data influences from pre-trained model without costly retraining, addressing privacy, copyright, and safety concerns. However, recent studies reveal a critical vulnerability: unlearned models rapidly recover "forgotten" knowledge through relearning attacks. This fragility raises serious security concerns, especially for open-weight models. In this work, we investigate the fundamental mechanism underlying this fragility from a representation geometry perspective. We discover that existing unlearning methods predominantly optimize along dominant components, leaving minor components largely unchanged. Critically, during relearning attacks, the modifications in these dominant components are easily reversed, enabling rapid knowledge recovery, whereas minor components exhibit stronger resistance to such reversal. We further provide a theoretical analysis that explains both observations from the spectral structure of representations. Building on this insight, we propose Minor Component Unlearning (MCU), a novel unlearning approach that explicitly targets minor components in representations. Extensive experiments on three datasets validate that, by concentrating unlearning effects in these inherently robust directions, our method achieves substantially improved resistance to relearning attacks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large language model, Unlearning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_XayROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work discusses unlearning information in large-language models and relearning attacks through the lens of the internal model representations on the 'forget set'. It provides an analysis into the components of the internal representations, and characterises how existing unlearning methods are weak to relearning attacks because the majority of the update they provide to models is concentrated in the dominant components which are easy to recover. Following this intuition, they present a method which explores using minor components for unlearning as they may be more robust. Their proposed method MCU, sits on top of existing unlearning methods and enhances their defence against relearning attacks. MCU is evaluated across model families and unlearning benchmarks, and each component is ablated strongly.

Overall, this paper presents an interesting analysis of LLM internal representations under unlearning but lacks sufficient experimental validation, statistical significance, and supporting analysis to justify the approach it presents. Given the concerns presented below, if the authors were to clarify I would be very open to increasing my score.

Critical strengths
  • The conceptual framework and exploration of dominant and minor components is well-structured and easy to follow, easily highlighting the need for a method which targets minor components.
  • The subspace projection method which restricts the update of various unlearning methods to the minor component subspace is well-defined, well-motivated, and easy to apply to various methods. The authors have done a good job of demonstrating the general machinery, and outlining how it can be applied.
  • The work presents strong ablations on subspace removal, model families, sensitivities, and component projection strategies.
Critical weaknesses
  • The analysis of principal components seems to be limited to a single model and data, with ablations of unlearning methods. This is interesting, but raises questions about the model and data. Furthermore, in the Appendix results I observe some settings in which the dominant components and recovery ratio do not support the claims made by the authors.
  • The main results show that MCU is complementary to existing unlearning algorithms with visible impact on forget quality as it does restrict the optimization. Because of this, and the restriction to minor components - it stands to reason that with MCU an unlearned model may be more viable to membership attacks which are a key consideration for unlearning. Can the authors expand on this?
  • The authors claim that MCU substantially improves on Relearning attack defence, yet in most comparative results visible in Table 1, the improvement is marginal and experimental results are not reported over multiple seeds with mean and standard deviation - which brings into question the statistical significance of these results. Furthermore, the most notable results are achieved with a combination of CIR and MCU (with a large percentage of the share coming from CIR alone) which is not mentioned by the text.
  • Additionally, the main result is that MCU + CIR gives "substantially lower post relearning accuracy and (∆) than all other configurations" yet this claim is not accurate as reflected in Table 5, (MLP Breaking + CIR + SAM outperforms), and in standalone comparisons SAM seemingly outperforms in many cases. This further calls into question what MCU is actually doing given the lack of statistically supported results.
  • Targeting minor components through the loss function is intuitive and interesting, but this changes the landscape, which fundamentally changes the problem that is being achieved through optimization. What seems to now be targeted is "what unlearning can be achieved through minor component modification only" rather than "can we unlearn in a way that is robust to relearning by only modifying minor components" which is what the work claims to address. This combined with uncertainties about the experiments, makes me call into question what the proposed method is achieving.
Questions for authors
  • What is the PCA of LLM representations of the presented methods with CIR, MCU? Does unlearning with CIR and/or MCU produce similar behaviour on unlearning change and/or relearning recovery? This could be a convincing result.
5 comments

We thank the reviewer for the constructive assessment of our representation analysis, the subspace intervention, and the ablations. We clarify the existing evidence for MCU's contribution, its relationship to CIR, and the scope of the evaluation below.


W1. Scope of the PCA analysis and exceptions to the recovery trend

Scope of the existing analysis. The paper evaluates the observations across objectives, sample sizes, and parameterizations:

EvidenceSettingsLocation
Main representation analysisLlama-3.1-8B on a representative forget datasetSection 3.1
Unlearning objectivesNPO, RMU, MLP Breaking, RMU + CIR, MLP Breaking + CIRAppendix L
PCA-fit and forget-set sizes25%, 50%, 75%, 100%Appendix M
ParameterizationFull fine-tuning; LoRA ranks 8, 16, 32, 64Appendix N

These experiments provide evidence that the observations extend beyond a single loss, fit sample size, or parameterization. Cross-model behavioral evidence is reported separately below.

How to interpret exceptions in the curves. Our mechanism concerns the relationship between the concentration of unlearning changes and their relative recoverability; it does not require recovery to decrease monotonically at every PC index or vanish in later directions. Some CIR curves show an initial decline followed by a flatter tail with substantial recovery. This qualifies the strength of the spectral separation in those settings, but does not by itself negate the greater recoverability of the leading directions. We will make this distinction explicit and avoid wording that suggests every minor component is immune to recovery.

Joint interpretation of change and recovery. The recovery ratio measures the fraction of an unlearning displacement that is reversed. It should be read alongside the change ratio: a large recovery fraction in a direction with little initial displacement has a different significance from recovery of a direction carrying much of the unlearning change. The paper reports both metrics for this reason. The theory explains the spectral tendency under its stated local approximations.

Behavioral generality across model families. Each entry below shows Relearn accuracy for MLP Breaking + CIR → MLP Breaking + CIR + MCU.

Paper reference: Appendix K, Table 8.

DatasetGemma2-9BQwen3-8B
WMDP-Cyber30.3 → 24.341.7 → 36.4
WMDP-Bio33.3 → 28.135.5 → 26.1
Years36.3 → 28.940.6 → 30.9

Adding MCU lowers Relearn accuracy across the reported settings, supporting transfer of the intervention across model families. This is behavioral evidence of generality; detailed cross-model recovery profiles are a separate mechanistic test.


W2. Forget quality and membership-inference risk

Effect on initial forgetting. The reported results do not support the premise that MCU systematically weakens forgetting. The following table compares MLP Breaking + CIR → MLP Breaking + CIR + MCU.

Paper reference: Table 1.

DatasetForget ↓Relearn ↓
WMDP-Cyber20.2 → 25.937.2 → 31.6
WMDP-Bio22.1 → 22.829.7 → 26.4
Years32.7 → 25.939.4 → 30.9

Initial Forget accuracy varies in both directions, while post-attack Relearn accuracy improves across these settings. On Cyber, MCU reaches approximately chance-level accuracy; the baseline's lower-than-chance accuracy alone does not establish more complete knowledge removal. The matched-Forget comparison below directly controls for different starting accuracies. We will emphasize absolute Relearn accuracy alongside the gap Δ.

Membership inference is a separate evaluation question. Restricting the representation-level objective does not, by itself, imply greater vulnerability to membership inference. Neither task accuracy nor the choice of subspace determines whether members can be distinguished from comparable nonmembers. Our current experiments evaluate knowledge recovery rather than membership inference, so we make no claim of improved membership privacy and have no evidence that MCU increases membership leakage.

Scope of the claim. A dedicated membership-inference evaluation would require member/nonmember splits with appropriate distribution and difficulty controls. We will distinguish this privacy question from the evaluated knowledge-recovery threat model. MCU's contribution here is empirical robustness against relearning, not a formal data-deletion guarantee.

W3. Independent contribution of MCU and statistical significance

Standalone contribution. The main results explicitly evaluate MCU both without CIR and on top of CIR, allowing their contributions to be distinguished. The comparisons without CIR give the following post-attack Relearn accuracies.

Paper reference: Table 1.

DatasetRMU → RMU + MCUMLP Breaking → MLP Breaking + MCU
WMDP-Cyber54.6 → 53.955.8 → 52.0
WMDP-Bio49.8 → 48.758.8 → 57.5
Years64.3 → 54.951.5 → 49.5

The reported comparisons favor MCU in absolute Relearn accuracy, but several differences are small.

Additional contribution on top of CIR. The paired comparisons below hold CIR fixed. Each entry reports Relearn accuracy for base objective + CIR → base objective + CIR + MCU:

DatasetRMU + CIR → + MCUMLP Breaking + CIR → + MCU
WMDP-Cyber50.5 → 43.337.2 → 31.6
WMDP-Bio47.6 → 39.529.7 → 26.4
Years37.4 → 33.739.4 → 30.9

These comparisons isolate the observed increment from adding MCU to an already effective CIR baseline. The combined result therefore includes a contribution from each method.

Why the two interventions differ. Appendix E provides a curvature-based interpretation of CIR. Under the local Fisher/GGN approximation and subspace-alignment assumptions, CIR's filtering of module inputs and output gradients retains parameter modes with lower relearning curvature. Their unlearning displacements are recovered more slowly under a limited attack budget:

\qquad \lambda_{ij}=\alpha_i\beta_j.$$ MCU instead specifies which representation directions enter the forgetting objective. These interventions operate in different spaces, connected by the model Jacobian; a filter on parameter modes does not generally impose the same objective as a transformation of representations before loss computation. This explains how both can affect recoverability while providing complementary constraints. The representation-change analysis and the paired comparisons above support this distinction empirically. The direction-wise recovery analysis and the curvature interpretation are also contributions beyond the combined method's performance. **Paper references:** Figure 3; Appendix E. **The evidence extends beyond small differences in individual cells.** The review's characterization of the gains as marginal does not account for the complete set of comparisons. The evidence supporting MCU includes: | Evidence | What it tests | Location | |---|---|---| | Standalone MCU across base objectives and datasets | Whether gains require CIR | Table 1 | | Adding MCU while keeping CIR present | The increment beyond the CIR baseline | Table 1 | | Nearly matched Forget accuracy and comparable WikiText loss | Whether gains arise from weaker initial forgetting | Table 6 | | Adaptive representation-restoration attack | Whether gains persist when the attack targets internal representations | Table 7 | | Additional model families | Whether the intervention transfers beyond the main model | Table 8 | The direction of improvement is consistent across these evaluations, and the CIR-controlled comparisons include substantial reductions in post-attack accuracy. The conclusion is supported by this repeated pattern across experimental conditions, rather than by treating a small difference in one cell as decisive. Variation in the magnitude of the gains does not explain away their consistency or the matched-Forget evidence. **Observed robustness and statistical testing are distinct claims.** The tables report empirical attack outcomes; we do not characterize them as results of a repeated-seed significance test. Cross-setting consistency is not a substitute for such a test, but it is relevant evidence for the method's effectiveness. Accordingly, the absence of seed-wise error bars does not by itself substantiate the suggestion that MCU contributes little beyond CIR or fails to improve robustness.

W4. SAM comparison and the claim of superiority over all configurations

The cited SAM result does not contradict MCU's incremental contribution. The comparison in the main-results discussion concerns the configurations in the main table. The complete results additionally include SAM combinations. The relevant question is whether MCU improves the corresponding unlearning pipeline and whether the broader comparison supports that contribution.

For MLP Breaking + CIR, the full robustness–utility comparison is shown below.

Paper references: Table 1 (main comparison); Table 5 (complete results).

DatasetMethodMMLU ↑WikiText ↓Relearn ↓Δ ↓
WMDP-CyberMLP Breaking + CIR65.01.00137.217.0
+ SAM65.31.00528.52.3
+ MCU64.71.00131.65.7
WMDP-BioMLP Breaking + CIR64.81.00129.77.6
+ SAM64.91.00231.45.5
+ MCU64.51.00126.43.6
YearsMLP Breaking + CIR64.61.01039.46.7
+ SAM64.31.02136.910.3
+ MCU63.81.01030.95.0

The complete comparison favors consistency over a single-task ranking. SAM achieves the lower attack accuracy on Cyber. MCU achieves lower Relearn and Δ on Bio and Years, while matching the baseline's WikiText ratio across the datasets. Moreover, MCU improves both Relearn and Δ relative to the same CIR baseline throughout; SAM raises Relearn accuracy on Bio and raises Δ on Years. The Cyber result therefore establishes a setting in which SAM is stronger, rather than showing that MCU's contribution is absent or generally inferior.

Standalone comparisons also require accounting for utility. For MLP Breaking without CIR, SAM has lower Relearn accuracy, accompanied by lower MMLU than MCU.

Paper reference: Table 5.

Dataset+ SAM MMLU ↑+ MCU MMLU ↑+ SAM Relearn ↓+ MCU Relearn ↓
WMDP-Cyber58.060.447.452.0
WMDP-Bio57.560.348.557.5
Years58.061.547.449.5

These are different robustness–utility operating points. They do not establish that SAM dominates MCU. MCU's contribution is the spectral explanation and the associated intervention that improves the paired baselines, including at matched initial forgetting. That claim is supported even when another method wins a particular configuration.

We will make the comparison set explicit in the main-results discussion and discuss the SAM trade-offs using the complete results.


W5. Does modifying the objective change the unlearning task?

The optimization objective changes; the evaluated unlearning task does not. Designing a surrogate that favors changes resistant to recovery is precisely the proposed approach to robust unlearning. MCU is evaluated on forgetting target knowledge, retaining general utility, and limiting held-out recovery under attack—not on minimizing the projected loss alone. Changing the surrogate therefore does not turn the evaluation into a different task.

Loss signal and actual changes. MCU directs the representation-level loss toward selected directions; it does not impose a hard constraint that every resulting representation change occurs only in those directions. Parameter sharing and the model Jacobian can couple changes across directions and layers. We will distinguish the targeted loss signal from the actual changes measured in the representation-change analysis.

Paper reference: Figure 3.

Matched-Forget evidence. The following comparison helps separate attack robustness from a change in initial forgetting strength.

Paper reference: Appendix I, Table 6.

DatasetMethodForgetRelearnWikiText ratio
WMDP-CyberMLP Breaking + CIR23.638.91.0001
+ MCU23.932.60.9991
WMDP-BioMLP Breaking + CIR23.229.61.0007
+ MCU22.825.41.0005

Here, MCU achieves lower Relearn accuracy at nearly identical Forget accuracy and comparable WikiText loss. This directly addresses whether the altered objective improves robustness at comparable initial forgetting strength. These comparisons match forgetting and utility rather than training steps; MCU requires more updates at these operating points.

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

This article argues that the existing methods for forgetting mainly optimize the dominant components while keeping the minor components largely unchanged. From a representation geometry perspective, it is found that during the relearning attack process, the modification of these dominant components can be easily reversed, thereby enabling rapid knowledge recovery, while the minor components exhibit stronger resistance to reverse operations. Therefore, a minor component forgetting (MCU) method is proposed. This method specifically targets the minor components in the representation for forgetting processing, thereby enhancing the resistance of the forgetting model in relearning.

Critical strengths
  1. The issue is significant and quite novel. The article explains "why existing LLM unlearning is so vulnerable to relearning attacks" from the perspective of representing geometry (dominant components/ minor components), and provides a testable mechanism.

  2. Clear and effective visualization. The paper provides a coherent visual narrative connecting representation spectrum, unlearning-induced changes, and relearning recovery. In particular, Figures 2 and 3 make the proposed dominant-versus-minor component hypothesis intuitive and show how MCU changes the allocation of representation updates.

Critical weaknesses
  1. The paper adopts the approach of using the MCU as the core. However, in the main result of Table 1, although the combination of CIR and MCU performed well, using only the MCU did not show a particularly significant improvement. In some indicators, it could not perform better than CIR, and it might even bring higher practical costs. Therefore, it cannot be determined whether MCU is an independent and effective robust unsupervised learning method, or merely an auxiliary modification, and only after updating the parameter space of CIR constraints does it become useful.

  2. In the ablation experiment, the mean direction was used as the 0th principal component in the MCU to prevent the global shared information from being restored. This in itself is not a major issue. However, other experiments (Middle/random) did not take this into account, which is unfair and cannot confirm whether it is the influence of different components or the mean direction itself.

Questions for authors
  1. Could the authors clarify whether MCU is intended to be an independently effective robust unlearning method, or primarily a complementary component for gradient-level methods such as CIR? In particular, can the authors provide a mechanistic explanation for why the representation-level projection of MCU alone is insufficient, and why CIR is needed to obtain the large robustness gains?

  2. The same as weakness 2, could the authors conduct fairer ablation experiments?

4 comments

We thank the reviewer for recognizing the significance of relearning robustness and the value of our representation-level analysis and visualizations. The review raises two important questions: how much MCU contributes independently of CIR, and whether the subspace ablation isolates the effect of principal-component selection. We address them below.

W1/Q1. Independent contribution of MCU and complementarity with CIR

MCU is a module for representation-based unlearning objectives, applicable with or without CIR. Here, “standalone MCU” means adding MCU to a base objective such as RMU or MLP Breaking without using CIR; MCU is not a complete unlearning objective on its own.

Standalone contribution. Table 1 includes the following comparisons without CIR. Each entry reports Relearn accuracy for the base objective → base objective + MCU:

DatasetRMU → RMU + MCUMLP Breaking → MLP Breaking + MCU
WMDP-Cyber54.6 → 53.955.8 → 52.0
WMDP-Bio49.8 → 48.758.8 → 57.5
Years64.3 → 54.951.5 → 49.5

MCU lowers Relearn accuracy across the reported settings without CIR. The magnitude varies across objectives and datasets, and standalone MCU does not dominate CIR in every setting. This establishes that MCU can contribute independently; its usefulness is not conditional on CIR being present. We also report utility alongside robustness, since some standalone settings involve utility trade-offs.

Additional contribution on top of CIR. Holding CIR fixed isolates the benefit of adding MCU. Each entry below reports Relearn accuracy for base objective + CIR → base objective + CIR + MCU:

DatasetRMU + CIR → + MCUMLP Breaking + CIR → + MCU
WMDP-Cyber50.5 → 43.337.2 → 31.6
WMDP-Bio47.6 → 39.529.7 → 26.4
Years37.4 → 33.739.4 → 30.9

CIR is an effective baseline, and these comparisons show that its robustness gains do not exhaust the benefit of MCU. The relevant test of complementarity is the improvement obtained by adding MCU while keeping CIR present.

Mechanistic distinction.

The two methods intervene at different points in optimization:

AspectCIRMCU
InterventionFilters module inputs and output gradients when forming parameter updatesTransforms representations before computing the unlearning loss
Geometric interpretationFavors parameter modes with lower relearning curvature under the local alignment assumptionsTargets minor representation components motivated by their observed recovery behavior
BasisUpdated during unlearningFitted to original-model forget-set representations and fixed during unlearning

Curvature-based explanation of CIR. Appendix E makes the connection explicit. For a linear module, CIR's filtered update takes the form

ΔWCIR=−ηuPg(gua⊤)Pa.\Delta W_{\mathrm{CIR}}=-\eta_u P_g(g_u a^\top)P_a.

Under a local Fisher/GGN approximation with Kronecker factors, a parameter mode has relearning curvature λij=αiβj\lambda_{ij}=\alpha_i\beta_j. If CIR's removed directions align with the leading eigendirections of those factors, its retained modes have smaller curvature. Their remaining unlearning displacement evolves approximately as

eij(Tr)≈e−ηrλijTreij(0).e_{ij}(T_r)\approx e^{-\eta_r\lambda_{ij}T_r}e_{ij}(0).

Thus, CIR can improve robustness by placing updates in parameter modes that relearning reverses more slowly. This interpretation depends on the local approximation and subspace alignment stated in Appendix E.

Why MCU adds something different. MCU makes the spectral allocation of the representation-level forgetting objective explicit. A parameter-space curvature filter and a representation-space loss transformation are different operations: the model Jacobian maps between the two spaces, so suppressing high-curvature parameter modes does not generally enforce MCU's representation-level objective. Conversely, targeting minor representation components does not impose CIR's filtering on both factors of a weight update. The methods can therefore reinforce each other without being equivalent.

Our contribution also includes the direction-wise analysis linking the concentration of unlearning changes to their subsequent recovery, and the curvature interpretation of CIR above. This explains the connection to CIR while identifying a distinct intervention supported by the paired comparisons.

Representation evidence. Appendix L includes the change/recovery analysis for RMU + CIR and MLP Breaking + CIR. Figure 3 compares MLP Breaking + CIR with its MCU variant: adding MCU shifts a greater share of representation changes toward later PC bins even when CIR is already present. This supports an effect beyond gradient filtering alone, consistent with complementarity. It does not establish that CIR is necessary for MCU to work.

Matched-Forget comparison and practical cost. Appendix I, Table 6 compares MLP Breaking + CIR with and without MCU at nearly matched Forget accuracy and comparable WikiText loss. Each entry shows baseline → + MCU:

DatasetForget ↓Relearn ↓Updates
WMDP-Cyber23.6 → 23.938.9 → 32.696 → 208
WMDP-Bio23.2 → 22.829.6 → 25.4154 → 231

W2/Q2. Fairness of the subspace ablation and control of the mean direction

The mean direction is part of the proposed spectral construction. MCU treats the mean as a shared dominant direction and removes it together with the leading centered PCs. Including the mean is therefore an intentional part of the method motivated by the representation analysis. Table 2 compares this construction with middle-PC and random-direction removal at the same prescribed removal dimension. Requiring each alternative to include the mean would define a different comparison: the effect of the remaining directions conditional on mean removal.

The paper already tests both subspace choice and mean handling. These questions are addressed by complementary ablations:

QuestionComparisonLocation
Does the proposed dominant-direction construction outperform arbitrary removal of the same prescribed dimension?MCU versus Middle and RandomTable 2
Does explicitly accounting for the mean affect robustness?MCU versus MCU w/o meanTable 3
Is an alternative spectral construction equally effective?Centered PCA with explicit mean handling versus uncentered SVDTable 3

Existing evidence on mean handling. Table 3 complements the rank-matched comparison by examining the treatment of the mean. All variants below use MLP Breaking + CIR; entries report Relearn accuracy:

DatasetCIR baselineMCUMCU w/o meanSVD-based
WMDP-Cyber37.231.631.731.1
WMDP-Bio29.726.435.429.2

On Cyber, removing centered dominant PCs without explicit mean removal already improves over the CIR baseline, and adding mean handling yields comparable Relearn accuracy. Thus, mean removal alone cannot explain the improvement in this setting. On Bio, explicit mean handling provides a stronger benefit. This supports the proposed design: both the shared mean and centered variation can matter, with their relative contributions depending on the dataset.

Why this does not invalidate the ablation. The suggestion that the benefit comes entirely from the mean is an alternative hypothesis, not a result established by the Middle/Random comparison. Table 2 evaluates the complete dominant-direction construction, while Table 3 explicitly varies mean handling. Together they support the proposed subspace design and the role of the mean within it. Our conclusion concerns that complete construction; it does not require attributing all gains exclusively to centered-PC selection. Holding the mean fixed across every strategy would further study interactions among these choices, rather than determine whether the evaluated MCU construction is effective.

Thank you for the careful and detailed response. My concern regarding the second weakness has been adequately addressed. However, I still believe that the improvement brought by MCU is less significant than that of CIR. Therefore, I would like to maintain my original rating.

Reviewer_4uPxOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes Minor Component Unlearning (MCU), a robust LLM unlearning method against relearning attacks. MCU is inspired by the observation that unlearning methods predominantly optimize along dominant components of MLP down projection activations, leaving minor components largely unchanged. During relearning attacks, the modifications in these dominant components are easily reversed, enabling rapid knowledge recovery. MCU can be applied to RMU and MLP breaking. Experiments are conducted on WMDP (Bio and Cyber) and Years. The paper also provides theoretical analysis that supports the observations.

Critical strengths

The presentation of the paper is generally clear, understandable, and easy to follow for me.

Critical weaknesses

However, I think the paper's current state does not meet the standard for ICLR acceptance for the following reasons.

Motivation, observations, and theory. The paper starts with observations alongside theorems that serve as the key motivations for the proposed method. I have critical concerns about their correctness and soundness, which I discuss below.

Observation 1 (Line 149-152)

Observation 1: LLM Representations are Concentrated in Dominant Components. Figure 2a shows the explained-variance ratio across principal components: the first few dominant components capture the overwhelming majority of the total variance, while the minor components form a long tail of small but non-negligible contributions.

  • Fig. 2a only shows the latent representations of the MLP down-projection matrix, but the stated observation about "LLM representations" is an overclaim.

  • By definition, PCA sorts principal components by explained variance. Therefore, observing that the first few principal components account for a large proportion of the total variance seems unsurprising.

  • The concentration of LLM representations in dominant components is already well known. Transformer activations are widely recognized to be anisotropic, with a few directions exhibiting large variance. This observation holds for general input text, not just forget inputs.

Observation 2 (Line 177-180) and Theorem 1 (Line 192-196)

Observation 2: Unlearning Predominantly Modifies Dominant Components. Figure 2b reports the Change Ratio (5) after applying GA: unlearning induces disproportionately large changes along the leading PCs, while minor components remain largely unchanged. The same pattern holds across unlearning losses (Appendix L).

Theorem 1 (Dominant-component concentration; explains Observation 2) etc.

  • Eq. 5's numerator defines the Change Ratio as the absolute projection of the representation change onto each principal direction, i.e., ∣⟨hu−ho,vk⟩∣|\langle \mathbf{h}_u - \mathbf{h}_o, \mathbf{v}_k \rangle|. As defined in the paper, minor components have small variances; their absolute changes will tend to be small. More concretely, consider RMU as an example: RMU pushes hu\mathbf{h}_u toward cuc\mathbf{u}. If RMU successfully unlearns the target knowledge, the change in representation can be approximated by cu−hoc\mathbf{u}-\mathbf{h}_o. Hence, by design, the projection of cu−hoc\mathbf{u}-\mathbf{h}_o onto vj\mathbf{v}_j scales with σk\sigma_k. Theorem 1 appears to derive an obvious result.

  • Furthermore, Theorem 1 intrinsically states that any loss that induces changes in latent representations will produce changes that scale with σk2\sigma_k^2. Observation 2, as explained by Theorem 1, appears to describe a general property of fine-tuning rather than a property specific to unlearning.

  • Lemma 1 in the Appendix also seems false in the case of RMU. Lemma 1 assumes that the noise term O(τ2)O(\tau^2) is small. However, in RMU, the coefficient of the target random vector is often large. In the original RMU paper, the authors select c=6.5c=6.5 (for Zephyr), which is more than twice the norm of the representation. Consequently, E[tt⊤]=c2uu⊤\mathbb{E}[\mathbf{t}\mathbf{t}^{\top}]=c^2\mathbf{u}\mathbf{u}^{\top} is a large rank-one matrix.

Overclaims. The proposed method is applied with RMU and MLP Break, but the paper repeatedly makes broader claims about "LLM unlearning." It is unclear how MCU can be applied to other unlearning methods. The current experiments do not sufficiently support the paper's broader claims about its general applicability to LLM unlearning.

I argue that the motivation and theoretical analysis of the proposed method are flawed and unconvincing, raising concerns about the method's correctness and soundness.

Questions for authors

Please see the Weaknesses.

3 comments

We thank the reviewer for the careful reading and detailed feedback. We appreciate the opportunity to clarify several important points. We believe that some concerns arise from an incomplete characterization of our contributions, while the concern regarding Lemma 1 highlights an assumption that deserves further clarification. We respond point by point below.

1. Observation 1: PCA ordering does not imply variance concentration

The reviewer argues that the concentration of variance in dominant principal components is unsurprising because PCA orders components by explained variance.

However, PCA ordering alone does not imply that a small number of components explain most of the total variance. PCA only guarantees

σ12≥σ22≥⋯≥σd2.\sigma_1^2\geq\sigma_2^2\geq\cdots\geq\sigma_d^2.

It places no constraint on the rate of spectral decay. For instance, an isotropic representation can have nearly uniform explained variance across all principal components despite the same ordering.

We agree that anisotropy in Transformer representations has been previously studied. We do not claim to be the first to identify anisotropy. Observation 1 establishes the representation geometry relevant to our subsequent analysis of unlearning and relearning.

This concern may stem from the wording of Observation 1. Here, “LLM representations” refers specifically to the internal MLP representations analyzed across Transformer layers, as described in Section 3.1 and Appendix F. Our focus on MLP modules is motivated by prior work identifying their important role in factual knowledge storage and retrieval [1,2], making them particularly relevant to understanding unlearning. We will make this scope and motivation more explicit in the wording.

[1] Geva et al. Transformer Feed-Forward Layers Are Key-Value Memories. EMNLP, 2021.

[2] Meng et al. Locating and Editing Factual Associations in GPT. NeurIPS, 2022.

2. Observation 2: Small original variance does not automatically imply small representation changes

The reviewer argues that minor components naturally exhibit smaller absolute changes because their original variances are smaller.

We believe this reasoning requires additional assumptions. In general, the variance of the original representation along a direction does not determine how strongly an optimization procedure can modify that direction.

For example, an optimization objective explicitly targeting a low-variance direction can induce a substantial displacement along that direction, irrespective of its original variance.

The relationship between original representation variance and unlearning-induced changes is therefore an empirical property of the optimization dynamics that requires explanation. This is precisely the purpose of Observation 2 and Theorem 1.

More importantly, Observation 2 is only one part of our central finding.

Our key motivation comes from the combination of:

  • Observation 2: Existing unlearning objectives predominantly modify dominant components.
  • Observation 3: Modifications along dominant components are recovered more rapidly during relearning, whereas minor-component modifications exhibit stronger resistance.

The reviewer focuses on Observation 2 without discussing Observation 3, which directly motivates the proposed MCU method.

Even if the concentration of unlearning updates partly reflects a general optimization phenomenon, the relationship between component-wise modifications and their subsequent recoverability remains central to understanding unlearning robustness.

3. Theorem 1: The contribution concerns the relationship between unlearning and relearning dynamics

The reviewer suggests that Theorem 1 may characterize a general property of fine-tuning rather than a phenomenon unique to unlearning.

We agree that the spectral structure of representations can influence optimization beyond unlearning. Our analysis does not require this spectral phenomenon to be exclusive to unlearning.

The central theoretical insight arises from combining Theorem 1 with Theorem 2.

Theorem 1 characterizes the concentration of unlearning-induced modifications along dominant directions. Theorem 2 then characterizes their recoverability under relearning attacks. Under the stated assumptions, dominant components with larger σk2\sigma_k^2 are recovered more rapidly, while minor components require a substantially larger relearning budget.

Together, these results explain a particular vulnerability of existing unlearning approaches: their modifications tend to concentrate in the directions that relearning attacks can recover most efficiently.

This connection between update concentration, component-wise recoverability, and robust unlearning motivates MCU's design. We believe that evaluating Theorem 1 in isolation overlooks this central contribution.

4. Lemma 1: The cited RMU coefficient differs from our experimental setting

We appreciate the reviewer's concern regarding the small-noise assumption in Lemma 1. However, the numerical argument relies on a steering coefficient from a different model configuration.

Experimental setting. The reviewer's numerical example uses the steering coefficient from the original RMU paper's Zephyr configuration. Our experiments use a smaller coefficient, as summarized below.

Under the centered-representation simplification used to analyze the RMU residual, the random-target contribution is

N=c2uu⊤,∥u∥2=1,N=c^2uu^\top,\qquad \|u\|_2=1,

and therefore

∥N∥op=c2.\|N\|_{\mathrm{op}}=c^2.

ConfigurationSteering coefficient ccRandom-target contribution ∥N∥op\|N\|_{\mathrm{op}}
Our experiments1.01.0
Zephyr setting cited by the reviewer6.542.25

The cited configuration therefore yields a much larger random-target contribution than our setting. This coefficient mismatch directly affects the reviewer's numerical objection to Lemma 1.

5. Generality: MCU provides a reusable mechanism for representation-based unlearning

MCU is designed for representation-based unlearning, and the paper explicitly states this scope. The Methodology section introduces MCU as a method that “explicitly redirects representation-based unlearning losses” toward the minor-component subspace. The Conclusion likewise states that “MCU is compatible with existing representation-based unlearning objectives.” These statements define the intended scope of the method.

Under “Overclaims,” the reviewer argues that the paper makes broader claims about “LLM unlearning” than the experiments support and questions how MCU applies to other methods. Our references to LLM unlearning describe the problem being studied; the method's applicability is explicitly specified in terms of representation-based objectives. Within this stated class, MCU provides a reusable mechanism that projects representations onto the minor-component subspace when computing the unlearning loss.

RMU and MLP Breaking have substantially different objectives: RMU steers representations toward a random target, while MLP Breaking disrupts their alignment with the original representations.

Demonstrating that the same minor-component projection mechanism improves both provides evidence that the proposed principle extends beyond a single loss formulation.

The reviewer's “Overclaims” criticism does not account for the scope explicitly stated in the Methodology and Conclusion. The results on RMU and MLP Breaking directly support the reuse of MCU across distinct objectives within that scope.

6. Empirical evidence for the effectiveness of MCU

Finally, we would like to emphasize that the review does not address the primary experimental evidence demonstrating MCU's effectiveness.

As shown in Table 1, adding MCU to RMU + CIR consistently reduces the relearning gap across all three datasets:

DatasetBase objective+ CIR, Δ ↓+ CIR + MCU, Δ ↓
WMDP-CyberRMU22.511.6
WMDP-BioRMU19.77.3
YearsRMU6.72.0
WMDP-CyberMLP Breaking17.05.7

These improvements are accompanied by broadly comparable model utility.

The benefits also extend to MLP Breaking, as illustrated by the Cyber comparison above.

Adaptive representation-based attack. Table 7 evaluates an attacker that explicitly attempts to restore the original model's internal representations:

DatasetMLP Breaking + CIR, Δ ↓MLP Breaking + CIR + MCU, Δ ↓
WMDP-Cyber16.62.9

These results provide direct empirical support for the proposed approach, including under a stronger attack that targets the representation-level modifications introduced by MCU.

We believe that these findings are important when assessing the overall soundness and contribution of the method, even while recognizing the limitations of the theoretical approximation.

Conclusion

We respectfully emphasize four points:

  1. The PCA and optimization arguments in the review do not invalidate our empirical observations. In particular, PCA ordering does not guarantee concentrated variance, and the review does not address the critical recoverability finding in Observation 3 and Theorem 2.
  2. The reviewer's numerical example uses a different RMU coefficient. The configuration comparison above shows a substantially smaller random-target contribution in our experiments, so the cited example does not establish a violation of Lemma 1's small-noise assumption in our setting.
  3. MCU's intended scope is representation-based unlearning, as explicitly stated in the Methodology and Conclusion. The results on RMU and MLP Breaking support its reuse across distinct objectives within this scope; the reviewer's “Overclaims” criticism does not account for this stated scope.
  4. MCU demonstrates substantial and consistent empirical robustness improvements, including under adaptive representation-based attacks, which are not addressed in the review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.