- Rating
- 2(+1)
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper introduces a two-stage LLM unlearning framework that constructs an audit signal through likelihood-gap optimization and suppresses sensitive outputs through inference-time sampling.
Critical strengths
-
The distinction between provider-side audit information and adversary-visible behavior is interesting.
-
The experimental comparison includes many established unlearning baselines and multiple model scales.
Critical weaknesses
-
The sampling formulation is inconsistent with the implementation. GapVeil claims a rejection-sampling distribution conditioned on an energy threshold. However, its adaptive token suppression changes subsequent proposal distributions, invalidating the exact sampling interpretation. The algorithm can also return a rejected candidate when its budget is exhausted.
-
Evaluation metrics and experimental configurations are problematic. The leakage metric assigns a perfect score to attacks with AUROC below 0.5, ignoring the possibility of reversing the attack decision. Moreover, Section 4.1 states that GapWiden is applied only to methods with audit AUROC below 0.95, whereas Appendix D.2 reports applying it to all eleven methods on MUSE. This protocol discrepancy requires clarification.
-
The conclusions are overstated. The method substantially decreases retain performance on MUSE-News, contradicting the broad claim of simultaneous improvement across evaluation dimensions. The experiments also lack convincing adaptive attacks against the inference-time filtering mechanism.
Questions for authors
-
What evidence shows that GapWiden's audit signal corresponds to actual data removal rather than deliberately engineered score separation?
-
What is the true output distribution of the adaptive GapVeil sampler?
-
How often does GapVeil exhaust its sampling budget and return a response that violates the acceptance threshold?
3 comments
the responses have addressed some of my concerns, so i have adjusted my scores,
Dear Reviewer kTxh,
Thanks a lot for your time and comments. We appreciate you for raising the score. We hope we have answered your questions. We would be happy to answer any other questions you might have.
Authors
We thank the reviewer for their comments, please find our response below.
How does Auditability validate Unlearning?
Thank you. Our goal is to make existing unlearning methods auditable, which we define as a statistical gap between the model’s behaviour on the forget and retain sets. We agree that such a gap alone does not rule out a model that still answers correctly when a question is reworded. We address this with two experiments:
What is the Output distribution of GapVeil?
Thank you for this question. Eq. 4 is the ideal distribution we wish to sample from. Sampling from it directly is impractical because of the cold-start problem described in Sec. 3.3. On forget queries, θ0 places most of its mass on the memorized answer, so independent proposals are repeatedly rejected. GapVeil therefore draws each new proposal from θ0 with the content tokens of earlier rejected drafts masked, so the proposal changes from one iteration to the next. In the revised draft (Appendix F, Prop. 1), we derive the exact output distribution of the sampler as implemented:
p^(y)=∑k=0KE[1[N≥k]qRk(y)]1[y∈A]+Pexhρ(y), where qRk is the masked proposal at iteration k, N is the index of the first accept, and ρ is the fallback used when the budget is exhausted. Conditional on acceptance at iteration k, the response is distributed as πRk∝cRkπ1[no masked token]. This is Eq. 4 restricted to responses that do not reuse content from rejected drafts, re-weighted by cRk. Each iteration is therefore an exact rejection sampler for its own target, and the output of GapVeil is a mixture of these targets and the fallback. The first iteration uses the unmasked proposal θ0, so its target is exactly Eq. 4, and it is selected with probability Z=pθ0(A). As a result, the TV(p^,π) is at most 1−Z, and the same bound holds for the distance between p^ and θ0 (Appendix F, Cor. F.1). On retain queries, where GapWiden keeps the gap near zero and Z is close to one, the deployed sampler is therefore close both to Eq. 4 and to the original model.
Exhaustion Rate
We report the exhaustion rate in Table-3 of general comment and add it to Appendix as well. We observe that an exhaustion rate of ~65%. The reason for this is that by design, θ0 and θu don't agree on forget set queries and hence, have a high energy (large gap G) becuase of which after several iterations, the gap doesn't go down to 0. Hence, after exhaustion we return the response with lowest energy across all steps. Note that this is by design and is expected. On the other hand, the exhaustion rate for retain set is very low, again because by design they are pinned to lower energy.
One sided MIA eval
Thanks a lot for pointing this out, in our earlier results, we had presented a one-sided leakage score in tables. However, after incorporating your comment, we have changed all the values as per Eq. 21 of the revised draft, which takes into account both sides of attack. Furthermore, we present the raw AUROC values in tables in Appendix for clarity. Broadly, our observations still hold that the proposed method improves the leakage score across all benchmarks: on average, from 0.79 to 0.96 on TOFU 1B, 0.76 to 0.96 on 3B, 0.77 to 0.95 on 8B, and 0.81 to 0.87 on MUSE. Further, we also see that the raw AUROC scores remain in the range of (0.4,0.6) avoiding attack from either sides.
GapWiden is applied to all methods on MUSE
We apply GapWiden to all methods on MUSE because we found that none of the methods were auditable on MUSE. This is in line with our protocol where we apply GapWiden only if the underlying method is not auditable.
Overstated claims
Thanks for the constructive critique, we would like to point that we have mentioned these limitations in detail in Appendix G including black-box and adaptive attacks. However, the focus of this work is to fundamentally question how unlearning should be approached?, where we emphasize on auditable unlearning which can be verified and presented as a certificate to deletion requests. We show such auditability is achievable without compromising downstream privacy, to our knowledge this is the first work to look at unlearning from this perspective.