acceptodds
Under review as a conference paper at ICLR 2027

Crush the Exposed Dark Side: Unlearning via Suppressing LLMs' Hazardous Output Tendency

Abstract

Large Language Models (LLMs) are widely deployed in real-world applications, yet their safety remains vulnerable to evolving jailbreak attacks. Existing defenses, including alignment and unlearning, typically rely on training data organized as question-answer pairs. As a result, their robustness is inherently bounded by the coverage of harmful prompts observed during training. Since natural language admits virtually unlimited reformulations, exhaustively covering the question space is infeasible in practice. To address this limitation, we propose Answer-Centric Unlearning (ACU), a framework that suppresses harmful answers exposed by the model itself. Specifically, ACU exposes latent harmful trajectories, iteratively suppresses them without conditioning on questions, and then trains on a retain set to restore normal utility. Experiments across four 8B backbones and one 32B model show that ACU improves robustness on eight additional jailbreak attacks while maintaining acceptable utility.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Unlearning, Language Models, Security

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 6 reviews, above 70% of the others.
Reviewer_14sCOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes Answer-Centric Unlearning (ACU), a method for suppressing harmful answers that is exposed by a model. The motivation is that model safety alignment is limited by coverage of prompts in the training data; ACU iteratively generates harmful trajectories from the model and then trains it to suppress these behaviors without conditioning on the query. The work first demonstrates that question-dependent unlearning can struggle to do well OOD while question-free (simply suppressing the output with no extra conditioning prompt) can do well OOD. The algorithm starts with a set of harmful prompts we want to suppress, and a set of benign answers to retain. The model is trained via an NLL loss to suppress the harmful ones, while still outputting the benign ones. Overall, they achieve a lower harmful-response rate across a variety of models compared to baselines on SORRY-Bench.

Critical strengths
  • The paper proposes a simple but seemingly effective method for suppressing unwanted responses in LLMs that gets strong results on the tested benchmarks
  • The paper contains thorough ablations on question-dependent vs question-free training for safety alignment and the effect of dynamically sampling new failures during training; measurements on how these affect OOD performance and other metrics are insightful and could be of use to the research community
Critical weaknesses
  • The novelty of the overall method appears to be limited, the pipeline seems to be largely well-known techniques like using on-policy rollouts whose benefits are already well-understood
  • Some simpler baselines that would strengthen the claims of the paper are missing such as:
    • Doing a simpler on-policy RL-based approach with a safety reward under a similar budget
    • Doing a simple SFT approach where the model is trained to answer with refusal under every adversarial prompt
Questions for authors
  • Could the authors provide results for leave-one-out experiments on other baselines? Table 2 only reports results on ACU, so its unclear whether the method generalizes better than other methods; this would strengthen the utility of the method substantially
  • Could the authors provide Table 8, 9 on other models besides Llama 3 (including the results of baseline methods on other models, not just ACU) to validate results are consistent across models?
4 comments

We thank the reviewer for the suggestions on methodological contributions and experimental completeness. We supplement the evaluation with simple safety-training baselines and category-holdout comparisons, and clarify the cross-model utility and over-refusal evaluations.

W1. ACU’s contribution beyond on-policy training

We agree that the benefit of on-policy sampling is well known, and our target-source control is consistent with it (4.35% HR with the model's own targets vs. 6.53% with another model's). The problem ACU addresses is that on-policy sampling from an aligned model finds very little to unlearn: exposure raises the number of accepted harmful-answer occurrences in the first collection pass from about 1,815 to 22,367. In the factorial comparison, refresh alone has little effect (14.48% → 14.51% without exposure) but helps after exposure (6.09% → 3.58%). The gain comes from combining exposure with on-policy collection, not from on-policy sampling alone.

Two further results point the same way. The on-policy RL baseline requested by the reviewer, which has no exposure step, reaches 10.94% HR (see W2). And the question-free objective matters independently of how targets are collected: with the answers fixed, QF beats QD on all four backbones, including at matched utility on Llama3 (5.35% vs. 8.30% HR at 70.76 vs. 70.82 utility).

W2. Comparisons with refusal SFT and on-policy safety RL

We add two baselines, both starting from the original aligned checkpoint and using the same adversarial probe pool with benign-task data:

  • Adversarial refusal SFT: pairs attack prompts with refusal responses and mixes them with benign question–answer examples.
  • On-policy safety RL: samples current-model responses and optimizes a safety reward, with benign-task training and KL regularization to the aligned checkpoint.

The table below specifies samples and updates, not matched GPU time. ACU incurs exposure, collection, suppression, and recovery costs; PPO incurs rollout generation and policy/value-update costs. Expanded recovery adds training cost.

PPO receives a reward of +1 for a safe response and −1 for a harmful one, as judged by a frozen Qwen3Guard, and is trained jointly with a cross-entropy loss on benign tasks. For both baselines, we searched learning rates {5e−6, 1e−5}, safety-to-benign ratios {1:1, 1:2}, and (for PPO) KL coefficients {0.01, 0.05, 0.1} on development data. The selected settings are:

ConfigurationRefusal SFTOn-policy PPO
Learning rate1e−55e−6
Safety-to-benign ratio1:11:1
KL coefficientNot applicable0.05
Training schedule1,000 updates, batch 161,000 rollout batches, one optimization epoch per batch
Sampling/minibatches16,000 example occurrences16 prompts × 4 responses per batch, minibatch 16
Total rolloutsNot applicable64,000 responses
Maximum answer length512 tokens512 tokens

The probe pool emphasizes rewritten harmful requests. Aligned Llama3 already has only 0.53% HR on the direct-request test panel, leaving little room for improvement on that panel. The broader evaluation measures transfer to jailbreak formats; the direct-request rate alone does not establish why these baselines improve only modestly.

MethodHR ↓Seven-task macro utility ↑
Aligned model11.82%72.16
Adversarial refusal SFT11.35%72.08
On-policy safety RL10.94%71.97
ACU, original recovery4.22%70.35
ACU, expanded recovery4.24%71.94

Both baselines keep utility near the aligned model but lower HR only slightly (11.82% → 11.35% and 10.94%). ACU lowers it to 4.24% at 71.94 utility. The expanded recovery set accounts for the utility: it raises ACU from 70.35 to 71.94 with comparable HR (4.22% vs. 4.24%), and lowers FRR on XSTest (7.43% → 6.97%) and on sensitive-benign prompts (9.07% → 8.53%).

Q1. Leave-one-category-out comparisons with other baselines

We agree, and have run the same protocol for three baselines on Llama3-8B: Safe Unlearning (answer-level unlearning), RMU (representation-level unlearning), and adversarial refusal SFT (refusal training). We chose one method per family; note that SafeLLM, the strongest baseline on the Llama3 main panel, is not among them.

The protocol is identical for all methods. For each held-out category, we remove its examples from all training and model-selection data (including paraphrases and records from other categories that contain the same intent), retrain, and test on the same held-out prompts: 100 intents × 5 responses × 3 seeds = 1,500 responses per category.

MethodWeapons HR ↓Physical harm HR ↓Biosecurity HR ↓Drugs HR ↓Cyber intrusion HR ↓Mean HR ↓
Aligned model8.60%12.33%9.87%11.13%13.47%11.08%
Safe Unlearning7.80%10.67%8.60%10.00%11.47%9.71%
RMU7.40%10.07%8.20%9.80%10.80%9.25%
Adversarial refusal SFT8.20%11.60%9.13%10.47%12.40%10.36%
ACU4.87%5.27%4.13%6.47%5.40%5.23%

ACU has the lowest HR in every held-out category. What the table does not show is that ACU loses less from holding a category out: its own mean rises from 2.53% (full data) to 5.23% (held out), and we do not have the baselines' full-data numbers on these category panels to compare against.

Q2. Utility and over-refusal on additional models

The paper currently covers the following:

  • All baselines: utility on Qwen3-32B (Table 20), and OR-Bench FRR on the four 7–8B backbones (Table 10, reproduced below).
  • ACU vs. aligned only: utility on the four 7–8B backbones (Table 34).

What is missing is baseline utility, XSTest, and sensitive-benign results on Qwen2.5, Qwen3-8B, and Ministral3. We will add these in the revision, with the same prompts, generation settings, and judge.

MethodLlama3Qwen2.5Qwen3Ministral3
Aligned model12.06%8.29%7.60%8.24%
Safe Unlearning9.25%11.63%10.74%11.38%
Eraser14.27%12.65%17.07%13.79%
RMU9.44%12.27%13.58%12.34%
Circuit Breakers9.38%9.41%9.66%9.17%
CKU7.99%11.77%10.93%12.92%
SafeLLM10.96%13.34%7.54%10.30%
ACU9.65%8.78%8.14%7.92%

Thank you for the additional experiments and clarifications. The authors have addressed most of my questions, although I still have concerns about the limited novelty of the method. Thus, I will maintain my already positive initial recommendation for weak acceptance.

We sincerely thank the reviewer for the thoughtful engagement and for maintaining a positive assessment of our work. We are glad that the additional experiments and clarifications addressed most of the concerns.

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

This paper proposes Answer-Centric Unlearning (ACU), which first temporarily exposes latent harmful behavior, iteratively collects harmful answers generated by the model itself, and suppresses these answers without conditioning on their eliciting questions. A small benign and refusal dataset is then used to recover utility. The experiments cover four 7–8B models and one 32B model, including fixed jailbreak templates, adaptive attacks, and weight-update attacks. The results show lower harmful-response rates than the aligned models and the evaluated safety baselines.

Critical strengths
  1. The evaluation is broad and the controlled comparisons are useful. The paper studies two model scales—four 7–8B models and one 32B model—under 28 fixed prompt conditions, two adaptive jailbreak attacks (PAIR and GCG), and three weight-update attacks. It also controls the harmful answers, update order, and recovery data when comparing question-free and question-dependent suppression.
  2. The recovery stage is reasonably well isolated. The recovery-only control remains at 11.11% HR, compared with 4.35% for fixed-target QF, showing that the final benign fine-tuning stage alone cannot reproduce the main suppression benefit.
  3. There is direct evidence that refreshing the target bank is useful. In the second and third collection passes, 26.63% and 16.91% of the harmful answers are assigned to clusters absent from earlier passes, while the held-out audit HR decreases from 52.16% to 9.59% and then 2.97%.
Critical weaknesses
  1. The foundational premises of the argument require verification; I believe this is a necessary addition to the motivation for the experiments. The paper’s core argument relies on the premise that "harmful outputs converge toward a limited number of generation trajectories." However, this premise is supported by only three citations, which actually substantiate related but distinct propositions. Furthermore, the paper provides no quantitative assessment of convergence within its own answer space. Consequently, the validity of these experimental premises remains to be empirically verified.
  2. The methodological novelty appears limited. ACU combines familiar components — malicious fine-tuning to elicit latent harmful behaviour, self-generated targets, negative-NLL unlearning, question-free conditioning, iterative target mining, and utility recovery — so the contribution is the integration and framing of these components rather than a new learning principle or algorithm. A simple method can still be a strong contribution if it yields a sufficiently surprising empirical finding, but as detailed in Weakness 1, that finding's premise is never directly validated.
Questions for authors
  1. Why six structurally distinct unlearning methods span only 7.32–11.02% mean HR across five backbones from four model families?Is this because the specific hyperparameters set for the baseline experiment led to these results?
  2. Is there direct evidence indicating that the space of harmful answers is smaller than the space of questions, or that it reaches coverage saturation more quickly?
1 comment

We thank the reviewer for the comments on our motivating premise and baseline comparisons. We clarify the cited evidence, quantify concentration in the collected targets, and highlight the controlled empirical findings behind ACU.

W1/Q2. Concentration and coverage of harmful answers

We agree: the three cited works support related claims, not the claim that the harmful-answer space is smaller than the question space. We will reword the passage so that each citation is attached to what it actually shows. Our own claim is narrower: many different questions lead to a more concentrated set of harmful answers.

To measure this directly, we took 10 harmful concepts, used 20 jailbreak prompts for each, and collected the successful harmful response to each prompt from the same exposed Llama3-8B checkpoint. We embedded prompts and responses with sentence-transformers/all-mpnet-base-v2 and computed the mean pairwise cosine similarity (i) among the prompts of a concept, (ii) among the responses of a concept, and (iii) between responses of different concepts. Row (iii) is a control: all responses come from the same model, so it controls for a uniformly high similarity associated with shared model style.

ComparisonMean cosine similarity ↑
Inputs within the same concept0.61
Harmful responses within the same concept0.83
Harmful responses across different concepts0.42

Responses to the same concept are more similar to one another (0.83) than the prompts that elicited them (0.61), and far more similar than responses to different concepts (0.42), consistent with concept-specific concentration beyond uniformly shared model style. Jailbreak wrappers and response style may still affect the embeddings. This supports the claim that varied prompts lead to similar answers. It does not show that the answer space saturates faster than the question space; that would require clustering and sampling curves on both sides, and we do not claim it.

Separately, Table 7 shows new-cluster fractions of 26.63% and 16.91% in the second and third passes, while independent audit HR falls from 52.16% to 9.59% and then 2.97%. These numbers show that refresh keeps finding new harmful modes as the model changes; they say nothing about the relative sizes of the question and answer spaces.

W2. Methodological novelty and empirical contributions

We agree that each component has precedent. What we contribute is the way they are combined and two findings that are not obvious in advance:

  • Training without the question transfers better than training with it. In Table 5, QD trained with eight questions per answer still has 6.62% HR on unseen attack expressions; QF, which never sees a harmful question, has 4.73%. Answers, retain data, update schedule, and recovery are identical.
  • Refresh benefits from exposure in the tested setting. Without exposure, refresh leaves QF HR essentially unchanged (14.48% → 14.51%); with exposure, it lowers HR from 6.09% to 3.58% (Table 6). The gain is therefore not just the known benefit of on-policy sampling.

The idea behind both findings is to turn malicious fine-tuning, normally an attack, into a way of finding what to unlearn: expose the model, collect the harmful answers it produces, suppress them without the question, and repeat on what remains. That the model's own answers matter is shown directly: with the recipient, target count, and budget fixed, self-generated targets give 4.35% HR, against 5.80% for targets from the aligned model and 6.53% for targets from another model.

Q1. Why are the baselines’ average HRs close, and could hyperparameters explain this?

The averages are close, but the methods behind them are not: they fail on different attacks. For example, on Llama3, Circuit Breakers beats Safe Unlearning on direct requests (0.17% vs. 0.77%) but loses on SORRY variants (6.42% vs. 3.55%).

On hyperparameters: all methods, including ACU, are tuned by the same rule (Appendix A.1). Configurations that degrade benign behavior too much are discarded, and the one with the lowest development HR is kept; test data plays no part. Each method is searched over its own hyperparameters. As an external check, we evaluated the released CB checkpoint for Llama3: under GCG, it is within two points of our reimplementation for both direct and jailbreak-wrapped inputs (see our response to Reviewer FBU1, W4). This provides an external check on CB specifically, not the other baselines.

We do not have a single explanation for why the averages fall in a narrow range. So that readers can check the role of hyperparameters themselves, we will add every baseline's final configuration and search range to the appendix.

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

This paper proposes Answer-Centric Unlearning (ACU), a defense against jailbreak attacks that suppresses harmful model-generated answers without conditioning the suppression objective on their eliciting questions. The framework first applies lightweight malicious fine-tuning to expose harmful responses, then iteratively collects and suppresses these responses while regularizing on benign examples. A final recovery stage uses benign and refusal examples to restore utility and refusal behavior.

Experiments cover four 7–8B backbones and one 32B model, comparing ACU against six defense baselines. Evaluation includes fixed jailbreak attacks, held-out harm categories, target-specific PAIR and GCG attacks, and fine-tuning-based attacks. Controlled experiments examine question conditioning, exposure, target refresh, capability preservation, over-refusal, and generation quality.

The results show substantially lower harmful-response rates than the evaluated baselines, with moderate utility costs and residual vulnerability to adaptive attacks.

Critical strengths
  1. The controlled experiments provide convincing evidence for the contribution of answer-only suppression. The matched-target QF/QD comparison fixes the starting checkpoint, harmful answers, benign records, update schedule, and recovery set. QF consistently achieves lower harmful-response rates across four backbones. Importantly, the Llama3 utility-matched comparison retains an advantage: QF achieves 5.35% HR versus 8.30% for QD at nearly identical aggregate utility. These controls make the contribution more persuasive than an end-to-end comparison alone.

  2. The robustness evaluation is broad and includes meaningful adaptive testing. ACU improves over the evaluated baselines across five backbones, and the paper additionally tests held-out harm categories, PAIR and GCG optimized separately against each defended checkpoint, and weight-update attacks. Reporting both response-level harmful rates and five-attempt success rates avoids equating low average harmfulness with reliable prevention. The accompanying over-refusal, generation-quality, and human judge audits further strengthen the empirical contribution.

Critical weaknesses
  1. The motivation overstates how far the method escapes question-coverage limitations. The suppression loss omits questions, but harmful-answer discovery still relies on a finite probe pool constructed from harmful intents and their rewrites. The held-out-category results show that coverage continues to affect effectiveness, while the local gradient analysis does not establish suppression of all harmful continuations. Appendix E.6 appropriately acknowledges these limitations. The evidence supports improved cross-question transfer, but not a general resolution of the coverage problem or certified knowledge removal. This matters because the stronger framing makes the conceptual contribution appear broader than the experiments establish.

  2. Utility recovery remains incomplete under the paper’s own evaluation criteria. On Llama3, aggregate utility decreases from 72.16 to 70.35 after the complete pipeline. Appendix H reports a recovered-minus-aligned difference of −1.81 percentage points, with a 95% interval of [−2.34, −1.28], exceeding the predefined acceptable loss of 0.5 points. The intervals for false refusal and degeneration also extend beyond their respective non-inferiority thresholds. The utility-matched control is valuable, but covers a fixed-target Llama3 comparison rather than the complete iterative pipeline across backbones. Consequently, the paper demonstrates a favorable safety–utility tradeoff more convincingly than full capability restoration.

Questions for authors

The following clarification requests could materially affect my assessment:

  1. Given that harmful-target collection remains probe-dependent, would you characterize the main contribution as improved cross-question transfer rather than independence from question coverage? What evidence most directly supports generalization beyond the behavior types represented in the probe pool?

  2. Can you clarify the main-text utility-preservation claim in light of the failed utility non-inferiority criterion in Appendix H? Are there complete iterative ACU checkpoints that approach the original model’s utility while retaining a substantial safety advantage?

  3. Appendix G shows that an aligned recipient supplied with exposed-source targets achieves 4.76% HR and 70.82 utility, compared with 4.35% HR and 70.12 utility for the exposed recipient, with an HR-difference interval including zero. Does this suggest that exposure should primarily be used for target generation? Has this alternative been evaluated with iterative refresh?

  4. Does the utility-matched QF advantage persist across additional backbones or for the complete iterative pipeline? The existing Llama3 control is persuasive, but broader evidence would strengthen the claim that the improvement is not primarily a consequence of capability loss.

2 comments

We thank the reviewer for recognizing the controlled evidence and for identifying where our claims need qualification. We address coverage, utility recovery, the role of exposure, and utility-matched transfer below.

W1/Q1. Cross-question transfer and probe coverage

Yes. We characterize ACU’s main contribution as improved cross-question transfer and reduced sensitivity to question coverage, rather than independence from coverage. We will revise the abstract and introduction to state this explicitly.

The most direct evidence for generalization beyond the probe pool is the leave-one-category-out experiment (Table 2; Appendix B.3), in which a whole category is removed from every training stage. For example, with all drug-related data removed, Llama3's HR on drug intents is 6.47%, against 11.13% for the aligned model; putting the category back lowers it to 2.53%. So ACU helps on categories absent from defense training, and helps more when they are covered.

W2/Q2. Utility recovery in the complete iterative pipeline

The original recovery set had only 50 examples (25 benign, 25 refusal). We replace the 25 benign examples with 400 drawn from four sources separate from our evaluation benchmarks: DeepMind Mathematics Dataset (mathematics), CodeSearchNet (programming), CLUTRR (relational reasoning), and Dolly-15k (question answering), 100 from each. The answers are generated by each backbone's own aligned checkpoint, before exposure. The 25 refusal pairs are unchanged, giving 425 examples. With this set, the remaining macro-utility loss is 0.22–0.66 points across the five backbones.

BackboneConditionMMLUHellaSwagARC-CBoolQPIQAWinoGrandeGSM8KAverage
Llama3-8BAligned64.1775.8055.6484.0578.6872.0074.7972.16
Llama3-8BOriginal recovery: 5062.3674.7153.4484.0775.2069.5373.1470.35
Llama3-8BExpanded recovery: 42564.1275.7055.5484.1278.1071.8374.1771.94
Qwen2.5-7BAligned65.3977.9158.2985.4180.0273.0576.0673.73
Qwen2.5-7BOriginal recovery: 5063.7776.4156.5284.3278.4571.5974.1972.18
Qwen2.5-7BExpanded recovery: 42565.0877.5357.4485.1679.3772.8875.2173.24
Qwen3-8BAligned69.0580.1661.4685.9180.2474.8480.0975.96
Qwen3-8BOriginal recovery: 5068.0878.6059.6385.1378.7373.2378.1574.51
Qwen3-8BExpanded recovery: 42568.9279.3460.0386.0479.9274.2078.6675.30
Ministral3-8BAligned67.1377.2158.4384.3679.3373.2877.0373.82
Ministral3-8BOriginal recovery: 5065.9175.3756.6683.6477.5471.8774.8472.26
Ministral3-8BExpanded recovery: 42566.7577.3057.8684.1178.6673.1976.8773.53
Qwen3-32BAligned74.8083.2067.4686.9481.7176.5087.6479.75
Qwen3-32BOriginal recovery: 5074.1282.5364.8685.4580.6576.1386.2778.57
Qwen3-32BExpanded recovery: 42574.8483.0766.1786.8881.7676.3687.2279.47

We also check that the larger recovery set retains the main-panel safety benefit. XSTest has 200 benign prompts and the sensitive-benign panel 50; five responses per prompt and three seeds give 3,000 and 750 decisions, respectively.

Llama3 metricAlignedOriginal recovery: 50 examplesExpanded recovery: 425 examples
Main-panel HR ↓11.82%4.22%4.24%
XSTest FRR ↓6.83%7.43%6.97%
Sensitive-benign FRR ↓8.27%9.07%8.53%

On Llama3, utility rises from 70.35 to 71.94 while main-panel HR stays comparable (4.22% vs. 4.24%), and FRR falls on both benign panels. The larger recovery set thus restores utility and reduces FRR while retaining the main-panel safety gain. On Llama3, the remaining utility gap to the aligned model is −0.22 points (paired 95% CI [−0.49, 0.05]). The lower bound is above the −0.5 threshold in Appendix H, so this checkpoint meets our non-inferiority criterion for utility.

Q3. Exposure for target generation versus recipient initialization

Appendix G supports the reviewer’s observation: an aligned recipient trained on exposed-source answers achieves strong safety performance with higher utility. Our reason for exposing the recipient is target availability during refresh. The first collection pass yields approximately 1,815 harmful-answer occurrences without exposure, against 22,367 with it. Exposure therefore supplies a high-yield starting point for sampling the recipient’s residual harmful outputs. Appendix G uses a fixed bank; we have not evaluated exposed-source targets followed by recipient-specific refresh from an aligned initialization. We will discuss that hybrid as an alternative design.

Q4. Utility-matched QF/QD comparisons across backbones

We extended the utility-matched comparison from Llama3 to all four 7–8B backbones. The protocol is unchanged. For each backbone, QF and QD share the exposed checkpoint, answer bank, retain records, update budget, and recovery. We search retain weights {0.25, 0.5, 1, 2} and learning rates {5e−6, 1e−5, 2e−5} on development data and, for each objective, keep the lowest-HR configuration whose utility is within ±0.25 points of a per-backbone target. No test data is used for selection. Values are mean ± SD over three seeds.

BackboneQF retain weightQD retain weightQF utility ↑QD utility ↑QF HR ↓QD HR ↓
Llama3-8B1.00.2570.76 ± 0.1070.82 ± 0.125.35% ± 0.18%8.30% ± 0.61%
Qwen2.5-7B1.00.2572.45 ± 0.1172.49 ± 0.135.42% ± 0.43%7.09% ± 0.55%
Qwen3-8B1.00.574.61 ± 0.0974.67 ± 0.124.58% ± 0.23%8.23% ± 0.48%
Ministral3-8B1.00.2572.57 ± 0.1272.65 ± 0.146.18% ± 0.42%8.42% ± 0.48%

At matched utility, QF has lower HR on all four backbones, by 1.67–3.65 points. This is still the fixed-bank setting, which isolates the training objective. We have not run a utility-matched QF/QD comparison for the full iterative pipeline; for the full pipeline, the relevant evidence is the expanded-recovery result in W2.

Reviewer_FBU1OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3(+1)
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Answer-Centric Unlearning (ACU), which suppresses harmful answer trajectories directly without question conditioning, followed by utility restoration on a retain set. As existing LLM safety defenses and training usually rely on question-answer pairs that is hard to cover all prompt reformulations, ACU tries to overcome remaining vulnerability to jailbreak attacks by systematically applying unlearning. Across 8B models and 32B model, ACU improves safety robustly against several jailbreak attacks with only a little trade-off of general capabilities. Still, ACU provides only behavioral suppression rather than knowledge deletion, and optimization based attacks retain non-trivial success when it compares to held-out harmful intents.

Critical strengths
  • The authors' idea of applying unlearning systematically to improve safety is good, as it is not just relying on a static question-answer pair datasets which is basic for those traditional safety trainings.
  • The logical structure of paper, starting from the motivation, mechanism to the method and experiment is easy to follow and quite valid.
  • Ablation studies including the differentiation between QF and QD are essential for distinguishing the actual role of each subpart of ACU. Analysis itself is also thorough and well-written.
Critical weaknesses
  1. Although the general results make sense, the general utility accuracy seems to drop a little bit eventually. That means, I think it would not be better one than the recent safety finetuning, representation altering methods such as RepBend [1] or ASGuard [2], which preserve general ability perfectly.
  2. The second subsection of Related Work, Representation control and model editing, is a little bit short and not enough.
  3. One of the most confusing part of this research is some terminologies. The authors use harmful-response rate, the HR, primarily on evaluation via GPT judge without providing explicit conceptual difference or justification for preferring HR over the standard Attack Success Rate (ASR) which is more widely adopted in AI safety papers and benchmark evaluations. And yet, on Page 6, the authors introduce intense level success within five attempts as ASR_5. If HR is used as the main metric, it is unclear why the intense-level metric is framed as ASR rather than HR. In addition, instead of adhering to the well-established concept of over-refusal, the paper newly introduced False-Refusal Rate (FRR). The terminological inconsistency hinders straightforward comparisons with existing methodologies and benchmarks that report standard ASR and over-refusal rate.
  4. There are concerns regarding the fidelity of baseline reproductions and hyperparameter setup. For instance, in Line 284 of Optimization-based attacks, the paper reports a score of 22.47 for Llama3 8B instruct against GCG when defended by Circuit Breaker (CB). However, Zou et al. [3] demonstrated that CB reduces GCG attack success to near zero levels, and Yousefpour et al. [1] reported reproduced scores in the range of 3.44% to 4.37%. These discrepancies make it unclear whether the baselines were properly configured and optimized. If these results come from the methodologically distinct between HR and ASR, the exact points of divergence must be explicitly stated. Furthermore, even if such methodological differences exist, the paper should provide a thorough explanation for relatively high values observed for ASR_5.

[1] Yousefpour, A., et al., "Representation Bending for Large Language Model Safety"

[2] Park, Y., et al., "ASGuard: Activation-Scaling Guard to Mitigate Targeted Jailbreaking Attack"

[3] Zou, A., et al., "Improving Alignment and Robustness with Circuit Breakers"

Questions for authors
  • Some of the details are confusing. For example, do the authors only use OR-Bench 80k, without toxic and hard-1k?
  • The paper is easy to read, but it needs more paper work. Figure 1 is intuitive but needs more caption. The whole citation is hard to distinguish, as it is not separated with main text by parentheses.
  • It would be helpful to include the specific evaluation prompt used for the LLM judge like GPT-5 in the Appendix or code repository to ensure reproducibility. I think it could be one of the source of the gap between HR and standard ASR, so please explain why baselines exhibit notably higher vulnerability under HR than in prior studies using ASR.
  • It seems strange for the FRR reduction in safety trained models in Table 9 and 10. In Table 10, baseline defense methods often report lower FRR than the basic, aligned model. What mechanism allows these defenses to reduce over refusal compared to the base aligned model?
  • In Table 9 Sens. Benigh, 9.07 of ACU is bolded over 8.53 of CB, even though a lower FRR is desirable. It would be helpful to adopt a single, uniform bolding convention across tables and state this criterion in the table captions.
  • Is there any justification for the specific validation constraint detailed in the Appendix, such as limiting FRR increase under 3 percentage points and NLL increase up to 0.15 nats/token?
  • For some baseline, evaluating against established open checkpoints, like GraySwanAI/Llama-3-8B-Instruct-RR [1] for Llama3 8B Circuit Breaker, would make the empirical comparison more compelling.

[1] https://huggingface.co/GraySwanAI/Llama-3-8B-Instruct-RR

8 comments

We thank the reviewer for the careful assessment and for recognizing the motivation, structure, and controlled ablations. We address the concerns about utility, baseline comparisons, metric terminology, evaluation protocols, and presentation below.

W1. Utility preservation and comparison with recent representation-based defenses

RepBend and ASGuard are relevant reference points for utility preservation. We have not run them under our protocol, so we make no head-to-head ranking. The recovery experiment below shows how broader recovery data reduces ACU’s utility loss.

The original recovery set had only 50 examples (25 benign, 25 refusal). We replace the 25 benign examples with 400 drawn from four sources separate from our evaluation benchmarks: DeepMind Mathematics Dataset (mathematics), CodeSearchNet (programming), CLUTRR (relational reasoning), and Dolly-15k (question answering), 100 from each. The answers are generated by each backbone's own aligned checkpoint, before exposure. The 25 refusal pairs are unchanged, giving 425 examples. With this set, the remaining macro-utility loss is 0.22–0.66 points across the five backbones.

BackboneConditionMMLUHellaSwagARC-CBoolQPIQAWinoGrandeGSM8KAverage
Llama3-8BAligned64.1775.8055.6484.0578.6872.0074.7972.16
Llama3-8BOriginal recovery: 5062.3674.7153.4484.0775.2069.5373.1470.35
Llama3-8BExpanded recovery: 42564.1275.7055.5484.1278.1071.8374.1771.94
Qwen2.5-7BAligned65.3977.9158.2985.4180.0273.0576.0673.73
Qwen2.5-7BOriginal recovery: 5063.7776.4156.5284.3278.4571.5974.1972.18
Qwen2.5-7BExpanded recovery: 42565.0877.5357.4485.1679.3772.8875.2173.24
Qwen3-8BAligned69.0580.1661.4685.9180.2474.8480.0975.96
Qwen3-8BOriginal recovery: 5068.0878.6059.6385.1378.7373.2378.1574.51
Qwen3-8BExpanded recovery: 42568.9279.3460.0386.0479.9274.2078.6675.30
Ministral3-8BAligned67.1377.2158.4384.3679.3373.2877.0373.82
Ministral3-8BOriginal recovery: 5065.9175.3756.6683.6477.5471.8774.8472.26
Ministral3-8BExpanded recovery: 42566.7577.3057.8684.1178.6673.1976.8773.53
Qwen3-32BAligned74.8083.2067.4686.9481.7176.5087.6479.75
Qwen3-32BOriginal recovery: 5074.1282.5364.8685.4580.6576.1386.2778.57
Qwen3-32BExpanded recovery: 42574.8483.0766.1786.8881.7676.3687.2279.47

We also check that the larger recovery set retains the main-panel safety benefit. XSTest has 200 benign prompts and the sensitive-benign panel 50; five responses per prompt and three seeds give 3,000 and 750 decisions, respectively.

Llama3 metricAlignedOriginal recovery: 50 examplesExpanded recovery: 425 examples
Main-panel HR ↓11.82%4.22%4.24%
XSTest FRR ↓6.83%7.43%6.97%
Sensitive-benign FRR ↓8.27%9.07%8.53%

On Llama3, utility rises from 70.35 to 71.94 while main-panel HR stays comparable (4.22% vs. 4.24%), and FRR falls on both benign panels. The larger recovery set thus restores utility and reduces FRR while retaining the main-panel safety gain. On Llama3, the remaining utility gap to the aligned model is −0.22 points (paired 95% CI [−0.49, 0.05]). The lower bound is above the −0.5 threshold in Appendix H, so this checkpoint meets our non-inferiority criterion for utility.

W3. HR, ASR, ASR5, and over-refusal terminology

We will rename HR to "response-level ASR" (the fraction of responses judged harmful) and FRR to "over-refusal rate", and define ASR5 as the fraction of intents with at least one harmful response in five attempts.

W4. CB performance under GCG in our setting

Attack-input construction is a major difference from the cited evaluations. In Table 23, GCG suffixes are optimized on top of SORRY-Bench jailbreak variants (e.g., role play), not direct requests. Appendix B.4 omitted this detail; we will add it. With direct requests, CB’s HR is approximately 4–5%, with the released checkpoint’s 4.10% falling within the cited 3.44–4.37% range.

Released-checkpoint comparison (Q7). To check the role of our CB reimplementation, we also evaluated the released GraySwanAI/Llama-3-8B-Instruct-RR checkpoint (Q7). Both checkpoints are attacked on the same 200 held-out intents, in two forms (direct and jailbreak-wrapped), with the same GCG budget (20-token suffix, 250 steps, 128 candidates per step), decoding settings, and judge. GCG is run separately for each checkpoint and each input form. As in the paper, we score five fresh responses to the final prompt per intent and seed. Search outputs are excluded.

CB checkpointAttack conditionHR ↓ASR5 ↓
our reimplementationDirect request + GCG5.10%12.67%
Released RR checkpointDirect request + GCG4.10%10.83%
our reimplementationJailbreak-wrapped request + GCG22.47%45.50%
Released RR checkpointJailbreak-wrapped request + GCG20.83%43.17%

Two results stand out. Our reimplementation and the released checkpoint differ by less than two points in both settings. For the same checkpoint, wrapping the request raises HR from approximately 4–5% to approximately 21–22%. This establishes the effect of input construction within our protocol; search, decoding, and judging differences also matter when comparing separate papers.

Supporting representation diagnostic. A related multi-turn phenomenon is reported by Bullwinkel et al. (2025), who show that in multi-turn Crescendo attacks, the conversation context shifts the representations of harmful responses. We separately examined single-turn attack formats, including their independently optimized suffixes. For 50 intents, we took direct and jailbreak-wrapped versions in different wordings, optimized a GCG suffix for each, and teacher-forced the same reference answer. We then measured the cosine similarity between the answer-token representations of two prompts (averaged over tokens, then intents) at the two blocks where our CB loss is applied.

ComparisonOur CB loss block 10Our CB loss block 21
Different wording, same attack format0.910.88
Same wording, different attack format0.630.50
Different wording and attack format0.620.49

Changing the wording yields high similarity (0.91 / 0.88); changing the attack format yields lower similarity (0.63 / 0.50). This complements the behavioral result above. The comparison has two limits: the compared prompts also differ in their GCG suffixes, and this is not a measurement of CB's rerouting against its reference model. Blocks 10 and 21 are the loss layers of our reimplementation, not of the released checkpoint.

Q1. Which OR-Bench subsets were evaluated?

We use 10,000 benign prompts from OR-Bench-80K. All of Hard-1K is among them; it is part of the 10,000, not a separate panel. The Toxic subset is not used. Response quality is rated on 500 OR-Bench-80K prompts, 50 from each of the ten categories. We will state this in the revision.

Q3. Evaluation prompt and judge reproducibility

Appendix A.4 gives the request-aware rubric: actionable assistance for the harmful intent is harmful; refusal, safe redirection, and non-actionable discussion are non-harmful. Appendix A.5 and Table 12 validate the final GPT-5 judge on 200 human-labeled responses, with 96.5% accuracy. Qwen3Guard, used for probing, is validated separately. We will release the exact GPT-5 message format and parsing rules. Every method uses the same final judge and rubric. W4 addresses the input-construction and checkpoint differences behind the GCG comparison with prior work. The HR definition and labels are unchanged.

Q4. Why do some defenses reduce over-refusal?

The decrease occurs almost only on Llama3, and on OR-Bench. Aligned Llama3 starts with the highest FRR (12.06%, vs. 7.60–8.29% for the other backbones; Table 10), and every baseline except Eraser lowers it. On the other three backbones, baselines raise FRR in all cases but one (SafeLLM on Qwen3). On XSTest, every defense raises Llama3's FRR (Table 9). We have not isolated the cause. Our hypothesis is that fine-tuning on benign retain data loosens Llama3's unusually conservative refusal behavior on OR-Bench prompts, but we have not tested this. Note also that our panel includes all of Hard-1K, so its absolute rates should be compared with other OR-Bench results using matching panel composition.

Q6. Justification of the FRR and NLL validation constraints

These thresholds are practical filters, not derived quantities. During hyperparameter selection, we first discard any configuration that raises FRR by more than 3 points or benign NLL by more than 0.15 nats/token, and then keep the remaining configuration with the lowest development HR. The thresholds were fixed before any test evaluation and apply equally to ACU and all baselines. They are separate from the non-inferiority criteria in Appendix H, which apply to the final checkpoint; we will make this distinction clear.

W2/Q2/Q5. Related work and presentation

We will cite and discuss the relevant work in greater detail, including RepBend and ASGuard, and improve the paper’s presentation. This includes expanding Figure 1’s caption, correcting citation formatting, and adopting a uniform, explicitly stated highlighting convention across tables. In particular, the highlighting of 9.07 over 8.53 in the sensitive-benign over-refusal column will be corrected.

Reviewer_XccCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Answer-Centric Unlearning (ACU) for improving LLM robustness to jailbreaks. ACU first uses harmful fine-tuning to expose model-native harmful responses, then alternates probing with suppression of the collected responses under an empty-user chat template, while retaining benign-answer likelihood. A final supervised recovery stage improves benign behavior. The evaluation spans four 7–8B backbones and Qwen3-32B, fixed jailbreaks, checkpoint-specific PAIR/GCG attacks, restricted adapter-update attacks, and utility/over-refusal measurements. The strongest evidence is the fixed-target QF/QD comparison and the development-selected utility-matched comparison, which support an empirical benefit from removing question conditioning. I view the contribution as a useful training recipe and controlled empirical finding, with narrower conceptual novelty than the answer-space framing suggests.

Critical strengths

1.The paper offers a useful training insight. Removing the triggering query from the forgetting loss can improve generalization to other attacks. This gives practical guidance for designing safety training data and objectives.

2.The paper experiment is well designed

Critical weaknesses

1.The novelty seems limited. Circuit Breakers already uses response-only training without harmful user queries. The contribution here seems mainly to be a useful training combination, rather than a new answer-space idea.

2.The refresh comparison is not fully convincing. Extra samples from the static arm are not added to its training bank. A static method that actually uses the full sampling budget might do better than the reported control.

3.Utility is not fully preserved. Llama3 loses 1.81 points in average utility, which does not meet the paper’s own non-inferiority criterion. (Table 44)

4.Repeated attacks still have a substantial success rate. ASR5 remains around 22%–30% under adaptive attacks. The improvement is useful, but the remaining failure rate limits its practical protection. It should be noted that most modern attack methods are both repeated and adaptive.

5.The mechanism analysis is still quite local. The gradient and neuron results give some clues, but do not yet explain well when the transfer should work and when it should fail.

Questions for authors

1.Did both the native and matched-target Circuit Breakers implementations preserve the original response-only input construction? Please provide the targeted layers, reference checkpoint, loss/coefficient schedule, retain-data treatment, and the final configurations and training budgets of the native baselines. How were the matched-target adaptations tuned to account for their different supervision scales?

2.What happens if the fixed-bank arm uses all 88,920 initial-checkpoint candidates to construct its training bank, with the same 1,000 updates, benign retention and recovery? Reporting HR, utility, and the resulting target coverage on one representative backbone would help establish whether refreshing the model-dependent target distribution adds value beyond a better-utilized static collection budget.

5 comments

We thank the reviewer for the careful assessment. Below we clarify the relationship to Circuit Breakers, add the full-budget static control requested by the reviewer, report a larger recovery set, and narrow our mechanism and robustness claims.

W1. Novelty and the relationship to Circuit Breakers

We agree: response-only training has precedent in Circuit Breakers, and we will say so explicitly. The two methods differ in what they do with those responses. CB reroutes hidden representations at selected layers away from those of a frozen reference model; ACU raises the NLL of the harmful answer itself. The larger difference is where the answers come from: CB uses a fixed, pre-built dataset, whereas ACU samples them from the model being defended, after exposure, and re-collects them as suppression changes what the model generates.

AspectCircuit BreakersACU
Harmful suppression objectiveRepresentation rerouting relative to a frozen referenceDirect suppression of harmful-answer likelihood
Harmful training targetsPreconstructed datasetAnswers sampled from the recipient model
Target acquisitionStatic collectionExposure followed by iterative probing
Target updatingFixed training dataRecollection from the evolving model
RetentionRepresentation matchingBenign-answer likelihood retention and final supervised recovery

We will add this comparison to the revision.

Q1. Circuit Breakers implementation and matched-target adaptation

Yes, both of our CB variants present harmful responses under an empty-user template. Note that the CB in Table 1 is our reimplementation: the released Llama3 code mixes question–answer and response-only examples, whereas ours uses response-only inputs throughout. We will document this difference.

Main-table CB

  • Start and reference: the aligned backbone; a frozen copy serves as the reference.
  • Loss layers: blocks ⌊L/3⌋\lfloor L/3\rfloor and ⌊2L/3⌋\lfloor 2L/3\rfloor (zero-indexed; 10 and 21 for Llama3).
  • Adapters: rank-8 LoRA (scaling 16, dropout 0.05) on the attention and MLP projections of all blocks up to the higher loss layer.
  • Losses: for harmful inputs, ReLU of the cosine similarity to the reference representation; for benign and refusal retain inputs, the distance to the reference, normalized by d\sqrt{d}. The two terms are weighted α(1−t)\alpha(1-t) and tt, with t=(u−1)/(U−1)t=(u-1)/(U-1) at update uu of UU.
  • Budget: 1,000 updates, each with eight harmful and eight retain records; answers capped at 512 tokens.
  • Selected settings: Llama3, Qwen2.5, Ministral3: α=2\alpha=2, lr 10−510^{-5}; Qwen3: α=4\alpha=4, lr 5×10−65\times10^{-6}; Qwen3-32B: α=2\alpha=2, lr 5×10−65\times10^{-6}.
  • Search space: α∈{1,2,4}\alpha\in\{1,2,4\}, lr ∈{5×10−6,10−5}\in\{5\times10^{-6},10^{-5}\}, updates ∈{500,1000}\in\{500,1000\}, selected on development data.

Matched-target adaptation. It differs from the above in four ways:

  • It starts from the common exposed checkpoint, which is also frozen as the reference.
  • It uses the same harmful-answer bank, benign records, batch order, 1,000 updates, and recovery as QF/QD.
  • LoRA is applied to all blocks, to match QF/QD (the loss layers are unchanged).
  • The separate refusal-retain data is removed, and the scheduled weights are replaced by the constant objective RH+0.5 DBR_H+0.5\,D_B.

It is therefore a comparison of objectives on matched data, not the full CB pipeline.

Supervision scale. Table 40 tunes the retain weight and learning rate of each objective on development data to reach comparable utility, with the bank, updates, and recovery fixed. At comparable utility (70.76 vs. 70.88), QF has 5.35% HR and the CB adaptation 8.46%.

W2/Q2. A static bank that fully uses all 88,920 candidates

We ran the control suggested by the reviewer. We freeze the initial exposed checkpoint, generate all 88,920 candidates from it, and add every judge-positive response to the static bank, with the same acceptance rule as before and repeated answers kept. Everything else matches the refresh arm: 1,000 updates, the same benign retention, and the same recovery stage. Using the full budget does help the static arm (HR 6.09 → 4.96), so the concern was justified. Refresh is still lower at 3.58, a gap of 1.38 points, at essentially the same utility.

ConditionCandidates contributing to bank constructionAccepted harmful occurrencesDistinct behavior clustersUtility ↑Test HR ↓
Original fixed bank29,640≈22,3671,21870.486.09 ± 0.44
Full-budget fixed bank88,920≈66,3841,48670.294.96 ± 0.37
Refresh88,920≈16,2351,66470.443.58 ± 0.41

Results are on the factorial evaluation panel (fixed 1,000 updates), so they are comparable with Table 6 but not with Table 1. Accepted counts include repeated answers. Clusters group answers by harmful goal, action, and outcome and ignore wording, as in Table 7. Refresh also covers more distinct behaviors (1,664 vs. 1,486 clusters) even though it accepts about a quarter as many harmful answers. So its benefit is not only better use of the sampling budget: following the model as it changes surfaces behaviors that the initial checkpoint does not produce.

W3. Utility preservation and the non-inferiority criterion

The original recovery set had only 50 examples (25 benign, 25 refusal). We replace the 25 benign examples with 400 drawn from four sources separate from our evaluation benchmarks: DeepMind Mathematics Dataset (mathematics), CodeSearchNet (programming), CLUTRR (relational reasoning), and Dolly-15k (question answering), 100 from each. The answers are generated by each backbone's own aligned checkpoint, before exposure. The 25 refusal pairs are unchanged, giving 425 examples. With this set, the remaining macro-utility loss is 0.22–0.66 points across the five backbones.

BackboneConditionMMLUHellaSwagARC-CBoolQPIQAWinoGrandeGSM8KAverage
Llama3-8BAligned64.1775.8055.6484.0578.6872.0074.7972.16
Llama3-8BOriginal recovery: 5062.3674.7153.4484.0775.2069.5373.1470.35
Llama3-8BExpanded recovery: 42564.1275.7055.5484.1278.1071.8374.1771.94
Qwen2.5-7BAligned65.3977.9158.2985.4180.0273.0576.0673.73
Qwen2.5-7BOriginal recovery: 5063.7776.4156.5284.3278.4571.5974.1972.18
Qwen2.5-7BExpanded recovery: 42565.0877.5357.4485.1679.3772.8875.2173.24
Qwen3-8BAligned69.0580.1661.4685.9180.2474.8480.0975.96
Qwen3-8BOriginal recovery: 5068.0878.6059.6385.1378.7373.2378.1574.51
Qwen3-8BExpanded recovery: 42568.9279.3460.0386.0479.9274.2078.6675.30
Ministral3-8BAligned67.1377.2158.4384.3679.3373.2877.0373.82
Ministral3-8BOriginal recovery: 5065.9175.3756.6683.6477.5471.8774.8472.26
Ministral3-8BExpanded recovery: 42566.7577.3057.8684.1178.6673.1976.8773.53
Qwen3-32BAligned74.8083.2067.4686.9481.7176.5087.6479.75
Qwen3-32BOriginal recovery: 5074.1282.5364.8685.4580.6576.1386.2778.57
Qwen3-32BExpanded recovery: 42574.8483.0766.1786.8881.7676.3687.2279.47

We also check that the larger recovery set retains the main-panel safety benefit. XSTest has 200 benign prompts and the sensitive-benign panel 50; five responses per prompt and three seeds give 3,000 and 750 decisions, respectively.

Llama3 metricAlignedOriginal recovery: 50 examplesExpanded recovery: 425 examples
Main-panel HR ↓11.82%4.22%4.24%
XSTest FRR ↓6.83%7.43%6.97%
Sensitive-benign FRR ↓8.27%9.07%8.53%

On Llama3, utility rises from 70.35 to 71.94 while main-panel HR stays comparable (4.22% vs. 4.24%), and FRR falls on both benign panels. The larger recovery set thus restores utility and reduces FRR while retaining the main-panel safety gain. On Llama3, the remaining utility gap to the aligned model is −0.22 points (paired 95% CI [−0.49, 0.05]). The lower bound is above the −0.5 threshold in Appendix H, so this checkpoint meets our non-inferiority criterion for utility.

W4. Repeated and adaptive attacks

We agree: repeated, adaptive attempts are the realistic threat, and ACU does not provide complete protection. It does lower the success rate of such attacks relative to the strongest baseline we evaluated:

Llama3HR (CB → ACU)ASR5 (CB → ACU)
PAIR19.13 → 10.4341.67 → 26.50
GCG22.47 → 13.6745.50 → 29.83

Qwen3 shows the same pattern (Table 23). We will state the residual ASR5 in the abstract and conclusion, and note that these numbers hold only for the attack budgets we tested.

W5. Scope of the mechanism analysis

We agree that the analysis is local and does not give a full account of when transfer succeeds. It does identify two conditions under which transfer is stronger:

Model-native targets. A step on the model’s own answers raises question-conditioned NLL by 0.38 nats/token on Llama3, versus 0.07 for answers from another model (Table 38, all four backbones agree).

Later answer positions. The overlap between answer-only and question-conditioned gradient neurons rises from 0.36 at the first token to 0.54 at 32 tokens for the bomb-making concept, and rises in the same way for five further concepts (Table 37).

So transfer is weakest at the start of an answer and for targets the model would not itself produce. We do not claim more than this: the behavioral gains rest on the generation experiments and ablations, not on the mechanism analysis. We will say so explicitly in the revision.

Reviewer_3ULxOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the problem that LLM unlearning defense effectiveness is limited by the coverage of harmful prompts seen during training, finds that the same model's harmful responses converge to narrower generation trajectories (with NLL and gradient-enrichment evidence in Section 2), proposes Answer-Centric Unlearning (ACU): exposure (lightweight malicious fine-tuning) → probing to collect the model's own harmful responses → iterative suppression without the questions → retain-set recovery, defines the answer-only suppression loss (pairing harmful responses with an empty user message and maximizing their negative log-likelihood), and conducts evaluations covering five backbone models, ten attack conditions, category holdout, PAIR/GCG optimization attacks, and weight-update attacks, demonstrating that ACU achieves the lowest average harmful-response rate on all five models (the 7–8B tier drops from 11.82–12.67% for aligned models to 3.88–4.46%, and Qwen3-32B drops from 11.75% to 2.78%).

Critical strengths
  1. The core idea is novel and empirically supported: The answer-centric perspective contrasts sharply with existing question-dependent defenses, and the observation that "the same model's harmful outputs converge to narrower trajectories" is supported at both the NLL and gradient-enrichment levels in Section 2.
  2. The ablation design is rigorous with sufficient control variables: The QF vs QD comparison fixes the same batch of answers, updates, and recovery; across four backbones, QF's HR is about 4 percentage points lower. The utility-matching control (70.76 vs 70.82) rules out the explanation that "QF merely reduces capability." The control performing recovery only on the aligned model (HR 11.11%) confirms that the suppression contribution is irreplaceable.
  3. The 2×2×2 ablation fully quantifies the contributions of the three components: The exposure + refresh + QF combination is optimal (HR 3.58%±0.41). Removing exposure raises it to 14.51%, fixing the target to 6.09%, and reverting to QD to 12.87%. Table 7 shows iterative collection keeps discovering new patterns (new clusters account for 26.63% in the second round), validating the necessity of refresh.
Critical weaknesses
  1. Optimization attacks still achieve significant success rates: Under PAIR, ACU's HR is 10.43% and the five-attempt success rate is 26.50%; under GCG, HR is 13.67% and the five-attempt success rate is 29.83%. After weight-update attacks, ACU's absolute values remain non-trivial: 14.38% after MFT, 25.86% after AvO, and 26.03% after TrojanPraise. The method's practical defense strength may be overestimated.
  2. The Abstract's claim of "acceptable utility" is somewhat optimistic: Llama3's macro average drops from 72.16 to 70.35, with PIQA falling from 78.68 to 75.20. XSTest FRR rises from 6.83 to 7.43, and sensitive benign from 8.27 to 9.07. Section 5 simultaneously admits recovery is partial and that answer-only suppression may affect benign sensitive discussions.
  3. Coverage and bias issues of the probe pool: The probe pool is obtained by rewriting 856 intents 8 times each, then filtered by the exposed model down to 5,928 entries. After category holdout, drug's HR rises from 2.53% to 6.47% (11.13% for the aligned model), showing the effect depends on probe coverage. The shared pool filtered by Llama may be biased toward a particular backbone; the paper does not verify the pool's uniformity across different models.
Questions for authors
  1. The Abstract describes "four 8B backbones," but Section 4.1 lists Llama-3-8B, Qwen2.5-7B-Instruct, Qwen3-8B, and Ministral-3-8B-Instruct. Strictly speaking, Qwen2.5-7B is not 8B. Can the wording be unified to avoid misleading readers?
  2. Does the exposure phase introduce new security risks? Exposure temporarily relaxes alignment by fine-tuning on 50 pairs of harmful QAs, producing intermediate checkpoints. Do these checkpoints and the collected harmful responses require controlled access? If an exposed model is leaked, does it pose a greater threat than the original model?
  3. Can QF's NLL advantage on OOD jailbreaks in Fig. 2 be translated into behavioral-level HR differences? Fig. 2 is a conceptual, controlled experiment; the behavioral comparison in Section 4.3.1 is QF vs QD with fixed answers, but the correspondence between the NLL trajectory under OOD attacks and the final HR for the full iterative ACU pipeline is not directly shown.
2 comments

We thank the reviewer for the careful assessment and for recognizing the answer-centric contribution, the matched-target comparisons, and the factorial ablations. In response, we ran three additional experiments: a larger recovery set, probe pools filtered by each backbone, and a check that the NLL advantage in Fig. 2 carries over to generated responses.

W1. Residual success under optimization-based and weight-update attacks

We agree that the residual attack success shows ACU does not provide complete protection. Nevertheless, ACU consistently outperforms the compared defenses under the evaluated attack budgets. On Llama3, it reduces PAIR/GCG HR from 19.13%/22.47% with Circuit Breakers to 10.43%/13.67%, and five-attempt success rates from 41.67%/45.50% to 26.50%/29.83%. ACU also outperforms the baselines under adapter-level weight-update attacks: ACU's HR is 14.38%, 25.86%, and 26.03% under MFT, AvO, and TrojanPraise, against 48.71%, 39.77%, and 30.08% for the best baseline in each case.

W2. “Acceptable utility” and the limited recovery stage

The original recovery set had only 50 examples (25 benign, 25 refusal). We replace the 25 benign examples with 400 drawn from four sources separate from our evaluation benchmarks: DeepMind Mathematics Dataset (mathematics), CodeSearchNet (programming), CLUTRR (relational reasoning), and Dolly-15k (question answering), 100 from each. The answers are generated by each backbone's own aligned checkpoint, before exposure. The 25 refusal pairs are unchanged, giving 425 examples. With this set, the remaining macro-utility loss is 0.22–0.66 points across the five backbones.

BackboneConditionMMLUHellaSwagARC-CBoolQPIQAWinoGrandeGSM8KAverage
Llama3-8BAligned64.1775.8055.6484.0578.6872.0074.7972.16
Llama3-8BOriginal recovery: 5062.3674.7153.4484.0775.2069.5373.1470.35
Llama3-8BExpanded recovery: 42564.1275.7055.5484.1278.1071.8374.1771.94
Qwen2.5-7BAligned65.3977.9158.2985.4180.0273.0576.0673.73
Qwen2.5-7BOriginal recovery: 5063.7776.4156.5284.3278.4571.5974.1972.18
Qwen2.5-7BExpanded recovery: 42565.0877.5357.4485.1679.3772.8875.2173.24
Qwen3-8BAligned69.0580.1661.4685.9180.2474.8480.0975.96
Qwen3-8BOriginal recovery: 5068.0878.6059.6385.1378.7373.2378.1574.51
Qwen3-8BExpanded recovery: 42568.9279.3460.0386.0479.9274.2078.6675.30
Ministral3-8BAligned67.1377.2158.4384.3679.3373.2877.0373.82
Ministral3-8BOriginal recovery: 5065.9175.3756.6683.6477.5471.8774.8472.26
Ministral3-8BExpanded recovery: 42566.7577.3057.8684.1178.6673.1976.8773.53
Qwen3-32BAligned74.8083.2067.4686.9481.7176.5087.6479.75
Qwen3-32BOriginal recovery: 5074.1282.5364.8685.4580.6576.1386.2778.57
Qwen3-32BExpanded recovery: 42574.8483.0766.1786.8881.7676.3687.2279.47

We also check that the larger recovery set retains the main-panel safety benefit. XSTest has 200 benign prompts and the sensitive-benign panel 50; five responses per prompt and three seeds give 3,000 and 750 decisions, respectively.

Llama3 metricAlignedOriginal recovery: 50 examplesExpanded recovery: 425 examples
Main-panel HR ↓11.82%4.22%4.24%
XSTest FRR ↓6.83%7.43%6.97%
Sensitive-benign FRR ↓8.27%9.07%8.53%

On Llama3, utility rises from 70.35 to 71.94 while main-panel HR stays comparable (4.22% vs. 4.24%), and FRR falls on both benign panels. The larger recovery set thus restores utility and reduces FRR while retaining the main-panel safety gain.

W3. Probe coverage and backbone-dependent filtering

To test whether filtering with Llama3 favors some backbones, we built five probe pools from the same candidate questions, each filtered by a different exposed model (Llama3-8B, Qwen2.5-7B, Qwen3-8B, Ministral3-8B, Qwen3-32B) using the original procedure. We then ran the full ACU pipeline on every backbone with every pool. The table reports main-panel HR.

Defended backboneLlama3-filteredQwen2.5-filteredQwen3-8B-filteredMinistral3-filteredQwen3-32B-filtered
Llama3-8B4.224.163.964.254.14
Qwen2.5-7B3.884.013.853.924.12
Qwen3-8B3.883.913.823.943.80
Ministral3-8B4.464.414.444.384.40
Qwen3-32B2.783.012.772.912.73

For every backbone, HR varies by at most 0.3 points across the five pools, and ACU remains below all baselines in each case. The choice of filtering model therefore has little effect in this comparison.

Q1. “Four 8B backbones” versus Qwen2.5-7B

We will consistently use “four 7–8B backbones and one 32B model” throughout the paper.

Q2. Security risks introduced by exposure

Yes. Exposure temporarily increases unsafe compliance, so exposed checkpoints, their adapters, and collected harmful responses require controlled access. We will state explicitly that these intermediate artifacts are excluded from unrestricted release and user-facing deployment, and distinguish them from the final recovered model.

Q3. Connecting Fig. 2’s OOD NLL advantage to behavioral HR

We agree that the NLL of a reference answer and the harmfulness of generated responses are different things. To connect them, we let the QF and QD checkpoints from Fig. 2 generate freely on the same OOD questions and scored the outputs. At the largest forget-set size:

ObjectiveReference-answer OOD NLL ↑Freely generated OOD HR (%) ↓
QD1.0721.40
QF1.1612.80

QF has both the higher OOD NLL and the lower OOD HR, so the likelihood result in Fig. 2 carries over to behavior in this comparison.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.