- 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 , 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 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
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 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 (’s 74% on Trees).
| Setting | Mask | Unmerge-CAV-Class | Best baseline | Kodge et al. | ||
|---|---|---|---|---|---|---|
| Trees, | 46.7 | 74.4 | 69.8 | 57.7 | 66.3 (SalUn) | 74.8 |
| Aquatic, | 46.4 | 75.8 | 70.3 | 56.5 | 66.9 (NegGrad+) | 76.6 |
| Vehicles, | 71.7 | 91.7 | 90.3 | 89.7 | 87.0 (SalUn) | 93.3 |
| Dogs, | 24.5 | 74.4 | 65.0 | 28.9 | 67.7 (Rand. Label) | 73.7 |
| Arthropods, | 22.9 | 82.5 | 67.9 | 39.3 | 75.4 (Rand. Label) | 78.9 |
| Trees, | 21.9 | 74.5 | 64.2 | 46.8 | 62.7 (SalUn) | 72.2 |
| Aquatic, | 20.4 | 66.5 | 66.3 | 41.5 | 61.0 (SalUn) | 67.0 |
| Vehicles, | 35.4 | 85.7 | 86.6 | 68.7 | 83.9 (NegTV) | 86.8 |
| Dogs, | 0.0 | 61.1 | 39.4 | 40.3 | 54.2 (Rand. Label) | 62.8 |
| Arthropods, | 0.0 | 69.5 | 28.8 | 49.1 | 55.5 (SalUn) | 71.0 |
- CAV-Class relearns less than the mask in 7/10 settings and is tied in two (Dogs : 40.3 2.5 vs 39.4 2.9; Vehicles : 89.7 1.3 vs 90.3 2.2); on Arthropods 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 it ends near (28.9 vs 24.5). Kodge et al., NegTV and NegMerge return to within 6 points of in every setting.
- The exception is Vehicles , where even relearns to 72% and every unlearned model returns to within 5 points of ; there SalUn is lowest (87.0 vs 89.7).
(c) Representation and probe. The feature gap is the mean relative gap to 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 .
| Setting | Feature gap : mask / Unmerge | Probe acc. (%): mask (= ) / Unmerge / |
|---|---|---|
| Trees, | 0.21 / 0.31 | 75.0 / 73.4 / 64.8 |
| Aquatic, | 0.76 / 1.59 | 75.0 / 73.4 / 63.8 |
| Vehicles, | 0.36 / 0.25 | 91.8 / 89.8 / 83.0 |
| Dogs, | 0.45 / 0.08 | 73.6 / 69.8 / 66.6 |
| Arthropods, | 0.27 / 0.06 | 81.2 / 78.2 / 71.0 |
| Trees, | 0.67 / 0.10 | 72.0 / 68.8 / 64.0 |
| Aquatic, | 1.44 / 0.38 | 65.0 / 58.0 / 53.8 |
| Vehicles, | 0.74 / 0.07 | 83.6 / 79.2 / 77.4 |
| Dogs, | 0.68 / 0.07 | 57.8 / 50.2 / 43.0 |
| Arthropods, | 0.35 / 0.26 | 66.0 / 54.6 / 60.6 |
- Unmerge’s features are closer to than the mask’s in 8/10 settings, including all and TinyImageNet settings (0.06-0.38 vs 0.27-1.44). The two exceptions are Trees and Aquatic , in line with App. F’s note on CIFAR-100 .
- The probe moves toward in every setting. It closes 14-29% of the gap on four settings and 40-71% elsewhere, and on Arthropods it goes past . As the reviewer notes, a decodable probe is not failure in itself ( 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.
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 grid. On the transfer tasks each method re-selects only its strength knob ( vs. ). We extended Kodge et al.’s grid until every selected configuration was interior, so a larger grid would not change their cells.
| Setting | Kodge et al., all layers (as published) | Kodge et al., our layers | Unmerge |
|---|---|---|---|
| Trees, / | 0.971 / 0.961 | 0.982 / 0.979 | 0.987 / 0.985 |
| Aquatic, / | 0.971 / 0.962 | 0.979 / 0.949 | 0.980 / 0.990 |
| Vehicles, / | 0.945 / 0.907 | 0.978 / 0.943 | 0.989 / 0.998 |
| Dogs, / | 0.842 / 0.837 | 0.935 / 0.853 | 0.990 / 0.915 |
| Arthropods, / | 0.813 / 0.767 | 0.924 / 0.787 | 0.996 / 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 .
- 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% (: 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 , with no leakage penalty and re-tuned over 15 values.
- Trees /: 0.974/0.975, against 0.987/0.985 for Unmerge and 0.982/0.979 for Kodge et al.’s own bases.
- Dogs /: 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.
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:
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) on forget-set images. MIA is the gap to WRT of a balanced, class-matched attack: forget-class training vs. test images, calibrated on forget-class data, chance 50%.
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 (WPT) and 0.89-0.99 (W0).
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 WFT’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):
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.