acceptodds
Under review as a conference paper at ICLR 2027

Unmerge: Efficient Machine Unlearning via Task Arithmetic

Abstract

Approximate machine unlearning seeks to remove the influence of a forget set from a trained model without full retraining. Existing gradient-based methods require data-dependent hyperparameter search, struggle when forget and retain knowledge are entangled, and offer little insight into where unlearning actually happens inside the network. We re-cast unlearning through the lens of task arithmetic: if finetuning produces a merged task vector that combines learning on forget and retain sets, unlearning is the inverse operation that subtracts a learned forget component to recover the retain task vector . The forget signal is concentrated: at every layer, forget activations lie in a subspace spanned by only a handful of dominant directions. We exploit this with a low-rank factorization , which is faithful up to a small tail-eigenvalue residual and bounds how far the correction can perturb retain. We then optimize three intuitive goals—match the merged vector inside the forget span, suppress leakage into the retain span, and bound the correction size—that provably bound forget leakage and retain damage in activation space. The resulting algorithm Unmerge is fast and powerful: on class-level unlearning with ResNet-50 on CIFAR-100 and Tiny ImageNet, it improves Tug-of-War by up to 24% over a baseline of comparable runtime and by up to 18% over stronger baselines that run slower, keeps membership-inference exposure at the level of retraining, and shrinks the feature-distribution gap to the retrained model, where relabeling methods leave forget features cleanly separable. Further studies show that Unmerge also applies to ViT-S/16 and scales to Llama-3.2-3B. The per-layer basis geometry that drives the algorithm also serves as a layerwise diagnostic for when and where unlearning becomes structurally hard.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Machine Unlearning, Model Merging/Unmerging

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_CBADOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Unmerge, a training-free method for concept-level machine unlearning. The method treats the parameter change caused by fine-tuning as a task vector and estimates a low-rank forget component from activation subspaces of the forget and retain sets. It then subtracts this component while penalizing overlap with retain directions. The method is evaluated mainly on class-level forgetting with ResNet-50 on CIFAR-100 and TinyImageNet, with additional experiments on ViT-S/16 and a preliminary Llama-3.2-3B study. The reported results show a favorable forget/retain trade-off and substantially lower runtime than several gradient-based baselines. The paper also studies layerwise subspace overlap, instance-level forgetting, and failure cases.

Critical strengths

1.The method explicitly models forget–retain interference. The retain-leakage term makes the edit weaker along directions shared with the retain set. This is more informative than applying a single global forgetting strength, and the term ablation shows that it is responsible for most of the improvement over plain projection.

2.The paper gives a useful and relatively honest analysis of scope. The authors test coherent and random instance-level requests and show that the method is primarily a concept-level editor. The layerwise diagnostics and failure analysis help clarify when the proposed approach is likely to work.

Critical weaknesses

1.The evaluation does not yet establish that Unmerge removes more training influence than simple output suppression. The appendix notes that masking the forget-class logits of the original model can already match retraining on ToW and the reported MIA gap. The paper does not include this as a complete quantitative baseline in the main comparisons. Feature-distance metrics provide additional information, but they do not directly show that the model is closer to the retrained model in its learned representation or training-data influence. The fact that a linear probe can still recover the forget classes is not, by itself, proof of failure, since a retrained model may also retain generalizable class information. Nevertheless, the current evidence is insufficient to determine the additional benefit of Unmerge over simple output masking.

2.The advantage over the closest subspace baseline and the practical tuning protocol are not sufficiently isolated. Unmerge changes several factors at once: the basis construction, soft leakage penalty, edited layers, and early-stopped solver. The Kodge et al. baseline is applied to all convolutional and linear layers, whereas Unmerge mainly edits deeper layers; the paper itself shows that restricting edits to deeper layers can substantially improve performance. Thus it is unclear how much of the gain comes from the proposed objective itself. In addition, the official configurations are selected using ToW, which requires the retrained model and test data. The proposed forget-accuracy-only rule helps for selecting α\alpha, but other hyperparameters still come from reference-dependent tuning, and the rule fails in some of the harder settings.

Questions for authors

1.Clarification about the evaluation: What are the complete results for the logit-masking baseline mentioned in Appendix F? Under the same protocol, how do logit masking, Unmerge, the unedited model, and the retrained model compare on output distributions, feature probes, and membership inference? Which metric still distinguishes Unmerge from simple output suppression?

2.Clarification about the comparison protocol: If Kodge et al. are allowed the same edited-layer selection, comparable data budget, and comparable hyperparameter-selection budget, does Unmerge retain a consistent advantage? Do the conclusions hold when all configurations are fixed without using the retrained model or the test set on the evaluation tasks?

3.Clarification about CAV-Class: Algorithm 3 separately centers the forget and retain pools. For a request containing only one class, the centered positive and negative pools each have zero total feature sum, which appears to make the ridge-regression target Z⊤yZ^\top y vanish. Does the implementation indeed produce a zero direction in this case, or is there an additional sampling or centering step not described in the paper? Please provide a single-class experiment.

5 comments
AuthorsOpenReview ↗

Response to Reviewer CBAD (1/4): Q1, complete logit-mask results (outputs and MIA)

We thank the reviewer for a careful and balanced review, and for recognizing the retain-leakage term and the scope analysis. We ran every comparison requested in Q1-Q2 on all 10 settings (5 forget tasks x 2 initializations), following the paper’s protocol, and Q3 on three new single-class requests.

Q1. Complete results for the logit mask. We agree that output metrics alone cannot establish the benefit. A logit mask is not itself an unlearning method:

  • it changes no weight and leaves all forget information in the network;
  • anyone with the weights removes it;
  • output filtering of a deleted class was proposed for the black-box setting (linear filtration [1]).

Because a mask never predicts a forget class, it scores near retraining on accuracy-based output metrics by construction, against any weight-editing method. Prior work therefore separates removal from suppression by probing representations [2, 3] and by relearning attacks [4, 5]. We report the mask as that control: its complete results are below and go into a new App. E.9 together with the relearning test, and Sec. 4 points to them.

(a) Output distribution and membership inference. KL is KL(p∥pRT)\mathrm{KL}(p \Vert p_{RT}) on forget-set images. MIA is the gap to WRTW_{RT} of a balanced, class-matched attack: forget-class training vs. test images, calibrated on forget-class data, chance 50%.

SettingToW: mask / UnmergeForget KL: WFTW_{FT} / mask / UnmergeMIA gap: WFTW_{FT} / mask / Unmerge
Trees, WPTW_{PT}0.995 / 0.98713.6 / 4.0 / 7.20.182 / 0.014 / 0.008
Aquatic, WPTW_{PT}0.994 / 0.98012.5 / 2.6 / 4.60.229 / 0.077 / 0.029
Vehicles, WPTW_{PT}0.998 / 0.98912.4 / 3.2 / 4.60.119 / 0.054 / 0.030
Dogs, WPTW_{PT}0.996 / 0.99015.2 / 2.2 / 3.20.228 / 0.025 / 0.052
Arthropods, WPTW_{PT}0.998 / 0.99616.8 / 1.8 / 3.00.203 / 0.020 / 0.027
Trees, W0W_00.989 / 0.98511.8 / 2.6 / 3.30.185 / 0.017 / 0.019
Aquatic, W0W_00.982 / 0.99011.3 / 2.1 / 2.70.274 / 0.075 / 0.137
Vehicles, W0W_00.983 / 0.99811.3 / 1.8 / 2.90.143 / 0.031 / 0.053
Dogs, W0W_00.999 / 0.91514.5 / 2.5 / 4.40.332 / 0.026 / 0.114
Arthropods, W0W_00.994 / 0.92115.3 / 2.3 / 5.10.294 / 0.044 / 0.113

Unmerge is the best variant for ToW, as in Tabs. 1-2. Elsewhere it is CAV-Class, our default when labels exist, fixed in advance. The retrain-to-retrain KL floor on Trees is 0.75-0.83 (WPTW_{PT}) and 0.89-0.99 (W0W_0).

Outputs and MIA do not separate the two. As App. F notes, a mask on the un-edited model matches retraining on ToW and the MIA gap: here it is at or above Unmerge on ToW in 8/10 settings and has the lower forget KL in every setting, and both cut WFTW_{FT}’s MIA exposure (0.12-0.33) to 0.01-0.14. This is expected, since all three metrics read only the outputs, which the mask edits directly.

What the weights still encode does separate them (Msg. 2):

  • Relearning: from five images per class, Unmerge-CAV-Class relearns the forget classes less than the mask in 7/10 settings (tied in two) and less than every baseline in 9/10.
  • Representation: its features are closer to WRTW_{RT} than the mask’s in 8/10 settings, including every W0W_0 and TinyImageNet setting.
  • Probe: it moves toward WRTW_{RT} in every setting, which a mask, leaving the features untouched, cannot do.

References

[1] Baumhauer, Thomas, Pascal Schöttle, and Matthias Zeppelzauer. “Machine Unlearning: Linear Filtration for Logit-Based Classifiers.” Machine Learning, vol. 111, 2022.

[2] Goel, Shashwat, et al. “Towards Adversarial Evaluations for Inexact Machine Unlearning.” arXiv preprint arXiv:2201.06640, 2022.

[3] Kim, Yongwoo, Sungmin Cha, and Donghyun Kim. “Are We Truly Forgetting? A Critical Re-examination of Machine Unlearning Evaluation Protocols.” Engineering Applications of Artificial Intelligence, vol. 167, 2026.

[4] Deeb, Aghyad, and Fabien Roger. “Do Unlearning Methods Remove Information from Language Model Weights?” arXiv preprint arXiv:2410.08827, 2024.

[5] Hu, Shengyuan, et al. “Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning.” ICLR, 2025.

AuthorsOpenReview ↗

Response (2/4): Q1, where Unmerge goes beyond a logit mask: relearning, representation, probe

Which metric distinguishes Unmerge from output suppression? Measurements of what the weights still encode: a relearning attack, the feature distribution and the probe. Output metrics do not, for any method (Msg. 1). The complete mask control and both tests go into App. E.9.

(b) Relearning: training-data influence. The reviewer asks for evidence beyond feature distances. A model that still encodes the forget data relearns it from a few examples much faster than one that never saw it [4, 5].

  • Protocol: fine-tune each model on 5 training images per forget class plus as many retain images, with 100 full-batch SGD steps (momentum 0.9), BatchNorm frozen, lr ∈{10−4,10−3}\in \lbrace 10^{-4}, 10^{-3} \rbrace and 3 seeds.
  • Reported: forget-class test accuracy at each model’s stronger lr.
  • Mask: made a weight edit by zeroing its classifier rows. Applied as an output filter, it is removed at step 0 (WFTW_{FT}’s 74% on Trees).
SettingWRTW_{RT}WFTW_{FT}MaskUnmerge-CAV-ClassBest baselineKodge et al.
Trees, WPTW_{PT}46.774.469.857.766.3 (SalUn)74.8
Aquatic, WPTW_{PT}46.475.870.356.566.9 (NegGrad+)76.6
Vehicles, WPTW_{PT}71.791.790.389.787.0 (SalUn)93.3
Dogs, WPTW_{PT}24.574.465.028.967.7 (Rand. Label)73.7
Arthropods, WPTW_{PT}22.982.567.939.375.4 (Rand. Label)78.9
Trees, W0W_021.974.564.246.862.7 (SalUn)72.2
Aquatic, W0W_020.466.566.341.561.0 (SalUn)67.0
Vehicles, W0W_035.485.786.668.783.9 (NegTV)86.8
Dogs, W0W_00.061.139.440.354.2 (Rand. Label)62.8
Arthropods, W0W_00.069.528.849.155.5 (SalUn)71.0
  • CAV-Class relearns less than the mask in 7/10 settings and is tied in two (Dogs W0W_0: 40.3 ±\pm 2.5 vs 39.4 ±\pm 2.9; Vehicles WPTW_{PT}: 89.7 ±\pm 1.3 vs 90.3 ±\pm 2.2); on Arthropods W0W_0 the zeroed-row mask relearns less (28.8 vs 49.1).
  • It relearns less than every baseline in 9/10 settings, by 6.5-38.8 points; on Dogs WPTW_{PT} it ends near WRTW_{RT} (28.9 vs 24.5). Kodge et al., NegTV and NegMerge return to within 6 points of WFTW_{FT} in every setting.
  • The exception is Vehicles WPTW_{PT}, where even WRTW_{RT} relearns to 72% and every unlearned model returns to within 5 points of WFTW_{FT}; there SalUn is lowest (87.0 vs 89.7).

(c) Representation and probe. The feature gap is the mean relative gap to WRTW_{RT} over four distribution distances on penultimate features (Sec. 4.4). The probe is a linear probe on frozen features, evaluated on forget-class test images. The mask’s features are those of WFTW_{FT}.

SettingFeature gap ↓\downarrow: mask / UnmergeProbe acc. (%): mask (= WFTW_{FT}) / Unmerge / WRTW_{RT}
Trees, WPTW_{PT}0.21 / 0.3175.0 / 73.4 / 64.8
Aquatic, WPTW_{PT}0.76 / 1.5975.0 / 73.4 / 63.8
Vehicles, WPTW_{PT}0.36 / 0.2591.8 / 89.8 / 83.0
Dogs, WPTW_{PT}0.45 / 0.0873.6 / 69.8 / 66.6
Arthropods, WPTW_{PT}0.27 / 0.0681.2 / 78.2 / 71.0
Trees, W0W_00.67 / 0.1072.0 / 68.8 / 64.0
Aquatic, W0W_01.44 / 0.3865.0 / 58.0 / 53.8
Vehicles, W0W_00.74 / 0.0783.6 / 79.2 / 77.4
Dogs, W0W_00.68 / 0.0757.8 / 50.2 / 43.0
Arthropods, W0W_00.35 / 0.2666.0 / 54.6 / 60.6
  • Unmerge’s features are closer to WRTW_{RT} than the mask’s in 8/10 settings, including all W0W_0 and TinyImageNet settings (0.06-0.38 vs 0.27-1.44). The two exceptions are Trees and Aquatic WPTW_{PT}, in line with App. F’s note on CIFAR-100 WPTW_{PT}.
  • The probe moves toward WRTW_{RT} in every setting. It closes 14-29% of the gap on four WPTW_{PT} settings and 40-71% elsewhere, and on Arthropods W0W_0 it goes past WRTW_{RT}. As the reviewer notes, a decodable probe is not failure in itself (WRTW_{RT} is decoded at 43-83%); a mask cannot move the probe at all.

References

[4] Deeb, Aghyad, and Fabien Roger. “Do Unlearning Methods Remove Information from Language Model Weights?” arXiv preprint arXiv:2410.08827, 2024.

[5] Hu, Shengyuan, et al. “Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning.” ICLR, 2025.

AuthorsOpenReview ↗

Response (3/4): W2/Q2a, Kodge et al. under matched conditions

Q2a. Kodge et al. with the same layers, data and selection budget.

  • Layers: Kodge et al. edit the same deep layers as Unmerge’s CAV variants.
  • Data: Kodge et al. now use class-stratified retain statistics, as in the official implementation. On TinyImageNet this raises them by 0.02-0.13 over Tab. 2; we update Tab. 2 and Tab. 13.
    • Using all images instead of 5,000 changes Kodge et al. by at most 0.002.
    • Unmerge with 5,000 class-stratified images stays within 0.012 of its reported cell (12/12).
    • App. C will state both methods’ statistics precisely.
  • Selection: both are selected by the same ToW protocol as in Tabs. 1-2, Kodge et al. over their own (αr,αf)(\alpha_r,\alpha_f) grid. On the transfer tasks each method re-selects only its strength knob (α\alpha vs. αf\alpha_f). We extended Kodge et al.’s grid until every selected configuration was interior, so a larger grid would not change their cells.
SettingKodge et al., all layers (as published)Kodge et al., our layersUnmerge
Trees, WPTW_{PT} / W0W_00.971 / 0.9610.982 / 0.9790.987 / 0.985
Aquatic, WPTW_{PT} / W0W_00.971 / 0.9620.979 / 0.9490.980 / 0.990
Vehicles, WPTW_{PT} / W0W_00.945 / 0.9070.978 / 0.9430.989 / 0.998
Dogs, WPTW_{PT} / W0W_00.842 / 0.8370.935 / 0.8530.990 / 0.915
Arthropods, WPTW_{PT} / W0W_00.813 / 0.7670.924 / 0.7870.996 / 0.921
  • As the reviewer expects, deeper-only editing helps Kodge et al., but Unmerge leads in 10/10 settings: by 0.041-0.134 on six of them, and by 0.001-0.011 on Trees and Aquatic/Vehicles WPTW_{PT}.
  • On those four CIFAR settings both methods are near the ToW ceiling; relearning, which we ran on Trees, separates them. Kodge et al. at our layers return to 74.2% and 72.5% (WFTW_{FT}: 74.4/74.5); Unmerge-CAV-Class to 57.7% and 46.8%.

How much comes from the objective itself? We give Kodge et al.’s projection rule our bases at our layers. In our parameterization it is the hard edit B=XQF−(XQR)C⊤B = XQ_F - (XQ_R)C^\top, with no leakage penalty and α\alpha re-tuned over 15 values.

  • Trees WPTW_{PT}/W0W_0: 0.974/0.975, against 0.987/0.985 for Unmerge and 0.982/0.979 for Kodge et al.’s own bases.
  • Dogs WPTW_{PT}/W0W_0: 0.965/0.858, against 0.990/0.915 and 0.935/0.853.
  • With the rule fixed, swapping bases moves ToW by -0.008 to +0.030. Replacing the projection with our objective adds +0.010 to +0.057 in every setting (best basis against best basis).

The objective is therefore the consistent source of the gain at matched layers, which is the retain-leakage term the reviewer highlights. It agrees with App. D.2, where that term alone recovers essentially all of the gain over plain projection.

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

The paper proposes Unmerge, a layerwise, low-rank weight-editing method for concept-level unlearning. It constructs forget and retain activation subspaces, parameterizes a correction as B Q_F^T, and balances removal of the task-vector response in the forget subspace against changes in the retain subspace. The resulting objective is a convex quadratic, although the main vision experiments apply an early-stopped Adam iterate rather than its exact minimizer. Experiments cover class-group removal on CIFAR-100 and TinyImageNet, a ViT extension, and a preliminary TOFU language-model study. The paper also investigates feature separation, membership inference, solver behavior, and failure on instance-level requests.

Critical strengths
  1. An efficient and interpretable editing procedure.

  2. The best variants achieve strong forgetting/retention trade-offs across both initialization regimes and multiple class groups.

  3. Substantial diagnostic and limitation analysis.

Critical weaknesses
  1. Stronger evidence of effective unlearning is needed. The results show that Unmerge reduces predictions for the target classes, but it is unclear how much this goes beyond simply blocking those predictions. The appendix acknowledges that a simple output mask already matches the main accuracy and privacy scores. Including this baseline in the main comparisons would make the method’s additional benefit clearer.

  2. Practical use still depends on substantial tuning. Although the paper emphasizes reduced tuning, the reported settings were selected using retrained-model and test-set results. These would generally be unavailable in practice. A fixed parameter-selection procedure tested on new unlearning requests would better support the practical claim.

  3. The connection between theory and implementation needs strengthening. The theory provides useful intuition, but it does not establish that the final model behaves like one retrained without the forget data. Performance also depends on stopping optimization early, while fully solving the proposed objective performs worse in some settings. The reliability of this stopping rule needs further evaluation.

  4. The improvement over closely related methods is not fully isolated. Prior work already uses activation subspaces for efficient unlearning. The comparisons differ in both the layers being edited and the way subspaces are constructed, making it difficult to identify the source of the gains. More controlled comparisons would clarify the contribution. [1] improves the forgetting–retention trade-off through sharpness-aware optimization. [2] studies beneficial training states and also identifies locally grokked LLM examples that are more amenable to unlearning. These results limit the novelty of the broad claim that favorable geometry makes unlearning easier. Unmerge's distinct contribution appears to be its particular low-rank weight edit and its computational efficiency; this should be distinguished clearly from the earlier mechanistic insights.

  5. The language-model evidence is preliminary. The LLM experiment uses one model, one forget split, and one seed. Unmerge also receives tuning while the comparison methods use default settings. Matched tuning budgets and repeated runs would provide stronger evidence of generalization.

[1] Sharpness-Aware Machine Unlearning

[2] Grokked Models are Better Unlearners

Questions for authors
  1. Can you add the acknowledged logit-mask control and retrain-to-retrain output-KL reference, and report a balanced, class-matched membership evaluation?

  2. What performance remains when the entire selection procedure is frozen without retraining or test-set access, including basis, rank, depth, and regularization?

  3. How sensitive is the actual early-stopped estimator beyond Trees, and can the Adam displacement argument be corrected? With separately centered forget/retain pools, what is the CAV-Class fallback for a single forget class, whose ridge-probe numerator is zero as written?

  4. How much of the improvement over Kodge et al. survives matched edited layers and common activation bases?

9 comments
AuthorsOpenReview ↗

Response to Reviewer ntLu (1/6): W1/Q1, output suppression vs. unlearning

We thank the reviewer for a detailed and constructive review. We ran every control and experiment requested in Q1-Q4. Unless a table lists fewer, they cover all 10 settings (5 forget tasks x 2 initializations) and follow the paper’s protocol: seed 1, last iterate, ToW selection, identical loaders. Code and logs will be released. The messages follow the order W1/Q1, W2/Q2, W3/Q3, W4/Q4 and W5.

W1/Q1. Does Unmerge go beyond blocking predictions? We agree this needs evidence beyond outputs, and we ran the mask on all 10 settings. A logit mask is output suppression, not unlearning, and it differs in kind from every method in Tabs. 1-2, Unmerge included:

  • What it changes. NegGrad+, Random Label and SalUn fine-tune the weights; NegTV, NegMerge, Kodge et al. and Unmerge edit them, changing what the network itself encodes. The mask changes no weight: it filters the forget logits at inference, and all forget information stays in the network.
  • How it is undone. Anyone with the weights drops the filter, and the forget classes return at once (WFTW_{FT}’s 74% on Trees).
  • Where it comes from. Output filtering of a deleted class was proposed for the black-box setting (linear filtration [3]), where the weights are never released.

Because it never predicts a forget class, the mask scores near retraining on accuracy-based output metrics by construction, against any weight-editing method. Prior work therefore separates removal from suppression by probing the representation [4, 5] and by relearning attacks [6, 7]. This is why Tabs. 1-2 compare unlearning methods, and why we report the mask as the reference point for these tests: a new App. E.9 subsection, “Suppression control”, gives its complete results with both tests, and Sec. 4 points to it. Reviewer CBAD asked about the same control, and both responses share these results.

SettingToW: mask / UnmergeFeature gap ↓\downarrow: mask / UnmergeProbe acc. (%): mask (= WFTW_{FT}) / Unmerge / WRTW_{RT}
Trees, WPTW_{PT}0.995 / 0.9870.21 / 0.3175.0 / 73.4 / 64.8
Aquatic, WPTW_{PT}0.994 / 0.9800.76 / 1.5975.0 / 73.4 / 63.8
Vehicles, WPTW_{PT}0.998 / 0.9890.36 / 0.2591.8 / 89.8 / 83.0
Dogs, WPTW_{PT}0.996 / 0.9900.45 / 0.0873.6 / 69.8 / 66.6
Arthropods, WPTW_{PT}0.998 / 0.9960.27 / 0.0681.2 / 78.2 / 71.0
Trees, W0W_00.989 / 0.9850.67 / 0.1072.0 / 68.8 / 64.0
Aquatic, W0W_00.982 / 0.9901.44 / 0.3865.0 / 58.0 / 53.8
Vehicles, W0W_00.983 / 0.9980.74 / 0.0783.6 / 79.2 / 77.4
Dogs, W0W_00.999 / 0.9150.68 / 0.0757.8 / 50.2 / 43.0
Arthropods, W0W_00.994 / 0.9210.35 / 0.2666.0 / 54.6 / 60.6

Unmerge is the best variant for ToW, as in Tabs. 1-2. Elsewhere it is CAV-Class, our default when labels exist, fixed in advance. The feature gap is the mean relative gap to WRTW_{RT} over four distribution distances on penultimate features (Sec. 4.4). The probe is a linear probe on frozen features, evaluated on forget-class test images. A mask’s features are WFTW_{FT}’s.

What the table shows:

  • Outputs: as App. F notes, a mask on the un-edited model matches retraining on ToW and the MIA gap. Here it is at or above Unmerge on ToW in 8/10 settings, with the lower output KL in every setting and the smaller MIA gap in 7/10 (Msg. 3), as expected for metrics that read only the outputs.
  • Representation: Unmerge’s features are closer to WRTW_{RT} than the mask’s in 8/10 settings, including every W0W_0 and TinyImageNet setting (0.06-0.38 vs 0.27-1.44). The two exceptions are Trees and Aquatic WPTW_{PT}, in line with App. F’s note on CIFAR-100 WPTW_{PT}.
  • Probe: it moves toward WRTW_{RT} in every setting, closing 14-29% of the WFTW_{FT}to-WRTW_{RT} gap on four WPTW_{PT} settings and 40-71% elsewhere; on Arthropods W0W_0 it goes past WRTW_{RT}. A mask cannot move the probe.

References

[3] Baumhauer, Thomas, Pascal Schöttle, and Matthias Zeppelzauer. “Machine Unlearning: Linear Filtration for Logit-Based Classifiers.” Machine Learning, vol. 111, 2022.

[4] Goel, Shashwat, et al. “Towards Adversarial Evaluations for Inexact Machine Unlearning.” arXiv preprint arXiv:2201.06640, 2022.

[5] Kim, Yongwoo, Sungmin Cha, and Donghyun Kim. “Are We Truly Forgetting? A Critical Re-examination of Machine Unlearning Evaluation Protocols.” Engineering Applications of Artificial Intelligence, vol. 167, 2026.

[6] Deeb, Aghyad, and Fabien Roger. “Do Unlearning Methods Remove Information from Language Model Weights?” arXiv preprint arXiv:2410.08827, 2024.

[7] Hu, Shengyuan, et al. “Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning.” ICLR, 2025.

AuthorsOpenReview ↗

Response (2/6): W1/Q1, relearning attack (where Unmerge goes beyond a mask) and KL reference

Relearning attack. This follows the held-out design of [6].

  • Protocol: the attacker fine-tunes each model on 5 training images per forget class plus as many retain images, with 100 full-batch SGD steps (momentum 0.9), BatchNorm frozen, lr ∈{10−4,10−3}\in \lbrace 10^{-4}, 10^{-3} \rbrace and 3 seeds.
  • Reported: forget-class test accuracy at each model’s stronger lr, as an adaptive attacker would choose.
  • Mask: we strengthen it into a weight edit by zeroing its classifier rows, since as a filter it is simply removed.
SettingWRTW_{RT}WFTW_{FT}Mask (zeroed rows)Unmerge-CAV-ClassBest baselineKodge et al.
Trees, WPTW_{PT}46.774.469.857.766.3 (SalUn)74.8
Aquatic, WPTW_{PT}46.475.870.356.566.9 (NegGrad+)76.6
Vehicles, WPTW_{PT}71.791.790.389.787.0 (SalUn)93.3
Dogs, WPTW_{PT}24.574.465.028.967.7 (Rand. Label)73.7
Arthropods, WPTW_{PT}22.982.567.939.375.4 (Rand. Label)78.9
Trees, W0W_021.974.564.246.862.7 (SalUn)72.2
Aquatic, W0W_020.466.566.341.561.0 (SalUn)67.0
Vehicles, W0W_035.485.786.668.783.9 (NegTV)86.8
Dogs, W0W_00.061.139.440.354.2 (Rand. Label)62.8
Arthropods, W0W_00.069.528.849.155.5 (SalUn)71.0
  • CAV-Class relearns less than even the strengthened mask in 7/10 settings and is tied in two (Dogs W0W_0: 40.3 ±\pm 2.5 vs 39.4 ±\pm 2.9; Vehicles WPTW_{PT}: 89.7 ±\pm 1.3 vs 90.3 ±\pm 2.2); on Arthropods W0W_0 the zeroed-row mask relearns less (28.8 vs 49.1).
  • It relearns less than every baseline in 9/10 settings, by 6.5-38.8 points. On Dogs WPTW_{PT} it ends near WRTW_{RT} (28.9 vs 24.5), while every baseline returns to 68-74%. Kodge et al., NegTV and NegMerge return to within 6 points of WFTW_{FT} in every setting.
  • The exception is Vehicles WPTW_{PT}, where even WRTW_{RT} relearns to 72% and every unlearned model returns to within 5 points of WFTW_{FT}; there SalUn is lowest (87.0 vs 89.7).

So outputs alone cannot tell any unlearning method from suppression. What the weights encode can: the mask leaves the features and the probe at WFTW_{FT}, and even as a weight edit it relearns more than Unmerge in 7/10 settings.

Retrain-to-retrain KL reference (Q1). Two retrained models also disagree on the forget set, so the KL values of Tab. 23 need a floor to be read. The floor compares two retrained models: different init seeds on W0W_0, different data orders on WPTW_{PT}. Values are forget-set KL(p∥pRT)\mathrm{KL}(p \Vert p_{RT}) on Trees; method rows are from Tab. 23, which lists every variant and split.

ModelW0W_0WPTW_{PT}
Retrain vs. retrain (floor)0.89-0.990.75-0.83
Un-edited WFTW_{FT}11.7513.64
NegGrad+4.895.89
Random Label5.206.62
SalUn5.216.62
Unmerge-CAV-KMeans2.895.29
Logit mask (suppression control)2.593.96
  • Among the unlearning methods of Tab. 23, Unmerge is closest to retraining. Unmerge-CAV-KMeans, the variant Sec. 4.5 reports, has the smallest forget-set KL in both regimes and closes 82% (W0W_0) and 65% (WPTW_{PT}) of the gap from WFTW_{FT} to the floor, against 61-64% and 55-60% for the gradient-based methods.
  • The mask’s lower KL (85% and 75% closed) is expected. On a forget image, a retrained model spreads its prediction over the retained classes, ranked by their similarity to the image. The mask keeps WFTW_{FT}’s retained-class logits, which carry the same ranking, and only removes the forget classes before renormalizing. This match is output-only, like its ToW and MIA scores (Msg. 1): its features stay those of WFTW_{FT}, and it relearns more than Unmerge in 7/10 settings (above).
  • Tab. 23 will gain this floor row.

References

[6] Deeb, Aghyad, and Fabien Roger. “Do Unlearning Methods Remove Information from Language Model Weights?” arXiv preprint arXiv:2410.08827, 2024.

AuthorsOpenReview ↗

Response (3/6): Q1, class-matched MIA; W2/Q2, selection without W_RT or test data

Balanced, class-matched membership inference (Q1). Protocol:

  • Members are forget-class training images and non-members are forget-class test images, matched per class and balanced (500 per side). Chance is 50%, and WRTW_{RT} scores 49-52%.
  • Attacker A is calibrated on half of the forget-class data (10 x 2-fold): threshold attacks on four signals, logistic regression, and an SVC on probabilities.
  • A shadow attacker calibrated on retain classes (the paper’s protocol, balanced) is at chance for every unlearned model and for WRTW_{RT}, and detects only WFTW_{FT} (61-82%). We therefore report attacker A.

Gap to WRTW_{RT} (lower is better):

SettingWFTW_{FT}MaskUnmerge-CAV-ClassKodge et al.Rand. LabelSalUn
Trees, WPTW_{PT}0.1820.0140.0080.0580.0460.033
Aquatic, WPTW_{PT}0.2290.0770.0290.1350.1090.123
Vehicles, WPTW_{PT}0.1190.0540.0300.0470.0650.069
Dogs, WPTW_{PT}0.2280.0250.0520.0680.1180.123
Arthropods, WPTW_{PT}0.2030.0200.0270.0500.1090.107
Trees, W0W_00.1850.0170.0190.0480.0600.039
Aquatic, W0W_00.2740.0750.1370.1740.1720.181
Vehicles, W0W_00.1430.0310.0530.0520.0740.062
Dogs, W0W_00.3320.0260.1140.1530.2030.201
Arthropods, W0W_00.2940.0440.1130.1280.1830.182
  • Unmerge reduces WFTW_{FT}’s exposure (0.12-0.33) to 0.008-0.137.
  • It is below Kodge et al. in 9/10 settings (tied on Vehicles W0W_0) and below Random Label and SalUn in 10/10.
  • A small gap must be read together with utility: a model that has lost the forget classes and much of the retained ones also looks like WRTW_{RT} to this attacker.
    • Among methods that keep ToW ≥0.9\ge 0.9, an Unmerge variant has the smallest gap in 9/10 settings (on Arthropods W0W_0 no baseline reaches ToW 0.9).
    • The exception is Vehicles WPTW_{PT}, where NegGrad+ scores 0.013 against our 0.028.
  • The mask is close by construction: its forget-class outputs carry almost no membership signal (Msg. 1).

W2/Q2. Selection without WRTW_{RT} or test data. Tabs. 1-2 select every method, ours and the baselines’, by ToW, as is common practice [8, 9]; one protocol for all methods keeps the comparison fair. We agree that a deployment usually has no retrained model, since avoiding retraining is the point of unlearning, so the practical claim needs a procedure without it. App. D.1 already gives a training-only rule for α\alpha: the smallest α\alpha with forget-set accuracy ≤1\le 1%. We now extend it to every knob and test it on new requests.

Rule. Use training-set accuracies only: among configurations with forget-set accuracy ≤1\le 1%, take the one with the highest retain-set accuracy (if none reaches 1%, the lowest forget-set accuracy). No WRTW_{RT} and no test data are involved.

  1. Tuning tasks (Trees, Dogs & Cats). We apply the rule to App. D.1’s grid, with depth extended by one step: basis, rank {8,…,128}\lbrace 8,\dots,128 \rbrace, depth {30,…,50}\lbrace 30,\dots,50 \rbrace, λ∈{5,20,50}\lambda\in \lbrace 5,20,50 \rbrace, γ∈{0.25,0.5}\gamma\in \lbrace 0.25,0.5 \rbrace and α\alpha.
    • It picks the ToW-optimal configuration in 10/12 variant x init cells.
    • It is never more than 0.002 below a reported cell.
  2. New requests, everything frozen. Every knob is fixed on the tuning task by the rule; on the new request only α\alpha is re-selected, by the same rule.
    • Three held-out superclass tasks (Aquatic, Vehicles, Arthropods; 6 settings): CAV-Class and CAV-KMeans match or exceed their reported, ToW-selected cells in 12/12. PCA is the weak case: within 0.001 on CIFAR WPTW_{PT}, but it overshoots on Aquatic W0W_0 and Arthropods (down to -0.12), as App. D.1 already reports for Arthropods.
    • Three new single-class requests (oak tree, baby, rocket; x 2 inits), frozen from Trees: CAV-Class reaches 0.992-0.997 (Msg. 4).
  3. Basis also chosen by the rule. In each of the 10 settings, the selected model is within 0.003 of the best reported Unmerge cell or above it.
    • Every baseline keeps its reported cell from Tabs. 1-2.
    • Unmerge’s reported margin over the best baseline exceeds 0.003 everywhere.
    • Hence Unmerge’s lead over every baseline in Tabs. 1-2 holds when Unmerge is selected without WRTW_{RT} or test data, and the selected model is also above Kodge et al. at our layers (Msg. 5) in all 10 settings.

We will add these results as a new paragraph and table in App. D.1, and cite them where Sec. 4 makes the practical claim.

References

[8] Fan, Chongyu, et al. “SalUn: Empowering Machine Unlearning via Gradient-Based Weight Saliency in Both Image Classification and Generation.” ICLR, 2024.

[9] Kurmanji, Meghdad, et al. “Towards Unbounded Machine Unlearning.” NeurIPS, 2023.

Reviewer_71QHOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper's key contribution is showing how to perform efficient, low-rank, geometrically controlled machine unlearning while providing both theoretical guarantees and insight into where forgetting occurs inside a neural network.

Critical strengths

The research of the task arithmatic is very strong for the unlearning and in the original paper that was introduced showed good promise. however, it comes with its own challenges and problems that the paper properly identified them.

Critical weaknesses

line 18 ) vague! It is not clear "what does faithful is associated with? what is faithful to what?" line 36) regulatory framework?! it sound ai generated! What is regulatory framework?

line 46) relabeling methods: so don't this method retrain/fine tune the model on the wrongly labeled samples?

line 45 - line 51) the description of the four methods, seems the way that AI would be generating the context.

line 52 ) fragility: please cite and reference papers for these statements - or show them with the experimental results. since this observation was discussed in the previous

line 149) why do you consider a "public" pre-trained checkpoint?

Questions for authors

see the previous section

2 comments
AuthorsOpenReview ↗

Response to Reviewer 71QH

We thank the reviewer for the positive assessment and for the close reading of our work. Below we address each point and quote the revised text (to be included in the revised manuscript).

L18, “faithful”: faithful to what? It means how exactly the factorization captures a task vector supported on the forget inputs. Revised abstract sentence:

“We exploit this with a low-rank factorization τF=BQF⊤\tau_F = B Q_F^\top, which captures any task vector supported on the forget inputs up to a residual bounded by the tail eigenvalues of the forget-activation covariance (Lemma 3.2), and which bounds how far the correction can perturb retain.”

L36, “regulatory frameworks”. We name the concrete provision instead.

“Data-protection law gives individuals a right to have their data erased (e.g., Art. 17 of the EU GDPR), and concerns about copyright, privacy and safety in foundation models call for deleting the influence of specific training data without retraining from scratch.”

L46, relabeling methods. Yes. Random Label fine-tunes the model on forget samples paired with random labels; in our runs this is interleaved with retain data. SalUn applies the same fine-tuning to the weights with the largest forget-loss saliency. We now say “fine-tune” explicitly.

L45-51, the four families. We rewrote the paragraph around the mechanism of each family:

“Most approximate methods fine-tune the trained model with a modified objective. Gradient-ascent methods such as NegGrad+ [1] maximize the loss on the forget set while minimizing it on retained data. Relabeling methods fine-tune on forget samples paired with random labels [2]; SalUn [3] restricts this fine-tuning to high-saliency weights. Teacher-student methods distill an ‘incompetent teacher’ on the forget set [4]. Fisher-based methods perturb, and saliency-based methods dampen, the weights estimated to encode the forget data [5, 6]. Each family trades forgetting against retained utility through its own hyperparameters.”

L52, fragility: citations or evidence.

  • The authors of SalUn [3] identify instability under varying hyperparameter choices as a core limitation of existing methods (their Fig. 1b).
  • The authors of NPO [7] show that gradient ascent drifts toward catastrophic collapse of model utility.
  • In our experiments, one fixed Random Label configuration leaves 9.24±5.099.24 \pm 5.09% forget accuracy across the three CIFAR-100 superclasses (Tab. 1, WPTW_{PT}).
  • Across three seeds of one task, SalUn’s and Random Label’s ToW vary by ±0.036\pm 0.036 and ±0.037\pm 0.037 on W0W_0, against ±0.008\pm 0.008 for Unmerge-CAV-Class (App. E.2, Tab. 14).

Revised sentence:

“(i) Fragility: gradient-ascent and relabeling methods are sensitive to their training hyperparameters, and the same configuration can under-forget on one request and degrade retained utility on another ([3, 7]; Tab. 1 and App. E.2), which forces per-request tuning.”

L149, why a “public” pretrained checkpoint? Unmerge forms the task vector τm=WFT−Winit\tau_m = W_{FT} - W_{init}, so the unlearner needs the weights the model was fine-tuned from.

  • “Public” meant only that this checkpoint is available to the unlearner. Here it is torchvision’s ImageNet-1k V1 ResNet-50, whose hash we release (App. D.4).
  • App. D.4 shows what happens otherwise: substituting a different pretrained checkpoint is catastrophic, while Winit=0W_{init} = 0 is a safe fallback when the true initialization is unknown.

We revise L149 to:

“a known pretrained checkpoint WPTW_{PT} (e.g., a publicly released one), or training from scratch (W0W_0)”

and add a pointer to App. D.4.

We hope these revisions resolve the reviewer’s concerns. Our responses to the other reviewers also add further clarifications and new results (e.g., a logit-mask control with relearning tests, and tuning without the retrained model); if the reviewer finds that our revisions and additional results strengthen the paper, we would be grateful if they would consider raising their confidence and rating.

References

[1] Kurmanji, Meghdad, et al. “Towards Unbounded Machine Unlearning.” NeurIPS, 2023.

[2] Graves, Laura, Vineel Nagisetty, and Vijay Ganesh. “Amnesiac Machine Learning.” AAAI, 2021.

[3] Fan, Chongyu, et al. “SalUn: Empowering Machine Unlearning via Gradient-Based Weight Saliency in Both Image Classification and Generation.” ICLR, 2024.

[4] Chundawat, Vikram S., et al. “Can Bad Teaching Induce Forgetting? Unlearning in Deep Networks Using an Incompetent Teacher.” AAAI, 2023.

[5] Golatkar, Aditya, Alessandro Achille, and Stefano Soatto. “Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networks.” CVPR, 2020.

[6] Foster, Jack, Stefan Schoepf, and Alexandra Brintrup. “Fast Machine Unlearning without Retraining through Selective Synaptic Dampening.” AAAI, 2024.

[7] Zhang, Ruiqi, et al. “Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning.” COLM, 2024.

Dear Reviewer,

The authors have kindly provided detailed responses to your comments. We would appreciate it if you could take a moment to review their rebuttal and share any further feedback or updates to your evaluation.

Thank you for your time and thoughtful consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: revised PDF uploaded, with the revisions in blue

We thank all reviewers for their constructive feedback. We have uploaded a revised PDF in which the revisions described in our responses are set in blue. New tables are numbered R1-R7, so every section, table and equation number cited in our responses still applies.

Summary of revisions. The reviewers who raised each point are in parentheses; our individual responses give the details.

  1. Unlearning vs. output suppression (ntLu W1/Q1, CBAD Q1). New App. E.9 reports a logit-mask control on all 10 settings (Tab. R6), together with two tests of what the weights still encode; a new paragraph in Sec. 4.5 summarizes it.
    • Under a relearning attack (Tab. R7), Unmerge-CAV-Class recovers the forget classes less than every baseline in 9/10 settings and less than the mask in 7/10 (tied in 2).
    • A linear probe on its features moves toward WRTW_{RT} in every setting, which a mask cannot do.
  2. Privacy and output fidelity (ntLu Q1). A balanced, class-matched membership-inference attack on all 10 settings is added to App. E.6 (Tab. R5), and a retrain-to-retrain KL floor to Tab. 23 (App. E.8, Sec. 4.5).
  3. Selection without WRTW_{RT} or test data (ntLu W2/Q2, CBAD Q2b). App. D.1 now selects every knob, the basis included, from training-set accuracies alone (Tab. R3), and Unmerge keeps its lead over every baseline in all 10 settings. Sec. 4.3 cites this result.
  4. Comparison with Kodge et al. (ntLu W4/Q4, CBAD Q2a). App. A runs Kodge et al. on the layers Unmerge edits and with common bases (Tab. R1). Unmerge leads in all 10 settings, and with common bases (Trees, Dogs & Cats) the gain comes from our objective. Kodge et al.'s TinyImageNet statistics are now class-stratified, as in the official implementation (Tab. 2, Tab. 13, Sec. 4.2, App. C).
  5. Theory and solver (ntLu W3/Q3a-b).
    • Sec. 3.3 states what the bounds show and what the experiments show.
    • The Adam displacement statement is corrected (Sec. 3.2, App. D.3).
    • App. D.3 extends the budget and exact-minimizer study to all settings (Tab. R4).
  6. Single-class requests (ntLu Q3c, CBAD Q3). Alg. 3 now handles a single forget class, with no change to any reported result, and App. B.1 adds single-class experiments (Tab. R2).
  7. Related work and scope (ntLu W4, W5).
    • Sec. 2 discusses the two recent works raised by Reviewer ntLu.
    • The abstract, Sec. 5, App. E.4 and App. F describe the language-model study as preliminary and list a matched-budget comparison as future work.
  8. Writing (71QH). The abstract (L18), the introduction (L36, L45-52) and Sec. 3.1 (L149) are revised as quoted in our response.

We thank the reviewers again for helping us improve the paper, and we would be glad to clarify anything further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.