- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes URGE, a loss-reweighting framework for machine unlearning that uses internal model representations to determine how much each sample or token should be forgotten. It combines two signals: one measuring how stable representations are under perturbations, and another tracking how much they change during unlearning. The method can be applied to existing unlearning approaches without requiring additional models or retain data for computing the weights. Experiments on image classification, diffusion models, and LLMs show generally improved forgetting–utility trade-offs compared to existing baselines.
Critical strengths
-
URGE can be integrated into several existing unlearning methods, which is great to make it broadly applicable.
-
The proposed weighting mechanism does not require retain data or additional models, which is a practical advantage.
-
The two proposed signals are simple and easy to understand.
-
I appreciate the evaluation across three different settings: image classification, diffusion models, and LLMs.
-
The paper includes extensive ablation studies that help assess the contribution of the different components.
Critical weaknesses
- Assumptions and applicability
The method relies on two assumptions that are not sufficiently justified: that representation invariance reflects the importance of forgetting a sample, and that changes in representations indicate successful forgetting. A model could substantially change its hidden representations while still retaining the memorized information, simply encoded differently.
Additionally, the proposed framework assumes a decomposable forgetting objective. It is unclear how URGE would apply to other gradient-based approaches, such as adversarial (i.e. AMUN), or bilevel methods. Can the authors clarify which methods URGE supports and how it could be extended to objectives that are not naturally decomposable into sample- or token-level losses? The framing of this needs to be improved in the manuscript.
- Image classification experiments
The classification results are strong, but the MIA evaluation is insufficient. I would like to see stronger membership-inference attacks, such as LiRA or RMIA. Additionally, the average gap in the literature is usually computed also with the fourth metric of privacy, the authors should include MIA in the overall evaluation and reporting an extended average gap to retraining. Accuracy is not enough for forgetting! Additionally, it would be useful to evaluate on a more challenging dataset, such as Tiny ImageNet, since CIFAR-10 and CIFAR-100 are relatively simple, low-resolution datasets.
- Diffusion experiments
For most baselines, forgetting accuracy is already close to 0%, making it difficult to assess whether URGE actually improves forgetting. The results mainly demonstrate better FID. I would like to see more challenging forgetting scenarios where the benefit of URGE is clearer for unlearning.
Furthermore, the diffusion experiments only evaluate class forgetting. Given the claims of general and fine-grained unlearning, I would like to see random instance-level forgetting, even with a signal forget set %. Finally, classification accuracy alone is not sufficient to establish successful forgetting. I would like to see diffusion-specific MIAs or other privacy metrics.
- LLM experiments
The LLM results are mixed, and I find some of the claims too strong given the experimental evidence. Retain extraction is often substantially worse than retraining, and adding URGE does not always improve performance. For example, on Llama-2-7B, GA+URGE considerably worsens forgetting extraction compared to GA. Can the authors explain this behavior?
Similarly, on MUSE, URGE improves some metrics while degrading others. I would like to see stronger and more consistent privacy evaluations across the LLM benchmarks too. These mixed results should also be discussed more clearly as a limitation.
- Validation of CRI and RS
I am not convinced that the proposed signals actually measure memorization or residual forgetting.
For CRI, high representation invariance could simply reflect robust or generalizable features rather than memorization. For example, a model might have memorized a rare fact about a fictional author, but replacing the original question with a generic prompt such as "Tell me about the author" could substantially change the representations without removing that information.
Similarly, in image classification, adding noise to model parameters might simply measure representation robustness to parameter perturbations. This is also quite different from changing the input prompt in LLMs. Why should these two approaches measure the same notion of forgetting importance?
More generally, the authors should validate CRI and RS against direct memorization and forgetting measurements and see if these actually correlate! Comparisons with simple controls, such as random weighting, token frequency, or uncertainty-based weighting, are also needed to establish whether the improvements come from the proposed signals or simply from changes in optimization.
- Approximate unlearning and guarantees
URGE does not provide certified guarantees, indistinguishability bounds, or proof that the target information has been removed. This is understandable for an approximate-unlearning method, but I would like the authors to acknowledge this limitation more explicitly, particularly when interpreting representation changes as evidence of forgetting.
Questions for authors
Please see the section on weaknesses and tackle the points raised. In general, the idea is interesting, but I am concerened as to how much the results actually validate the assumptions/hypothesis. I am not confident that the statements on how invariance relates to forgetting are trivial and need to see more results (see above) to make sure it is correct/valid. Finally, there are some limitations in the evaluation (see above) that I would like to see tackled, and very important, the work needs robust privacy metrics (RMIA, Lira, ...) as accuracy itself is not enough to claim unlearning.