acceptodds
Under review as a conference paper at ICLR 2027

ReSAT: Training LLMs Resistant against Refusal Ablation Attacks

Abstract

Open-weight language models allow downstream users to modify internal com- putations as well as prompts, creating a distinct challenge for preserving learned safeguards. We investigate whether safety training protects refusal and safety against the removal of multiple activation directions, rather than just one. We introduce Refusal Subspace Ablation, an attack that selects complementary di- rections and jointly removes their span throughout generation without updating model weights. Starting from one direction, it adds components outside the already selected span, choosing them by their effect under joint ablation. Our evaluations show that some current existing safety-trained models resist a selected single- direction attack yet remain vulnerable to joint ablation. To narrow this gap, we propose Refusal Subspace Ablation Training: mixed supervised fine-tuning fol- lowed by safety-training rounds that supervise safe responses under joint ablation and refresh candidate directions between rounds. We evaluate ReSAT on Llama- 3-8B, Mistral-7B, Qwen2.5-14B, against RSA and other refusal-ablation attacks, with jailbreak attacks and utility test. In key settings, ReSAT achieves lower attack success rates than most other baselines while preserving clean MMLU accuracy. By connecting adaptive attack evaluation with safety training, our work offers a practical path toward more robust safeguards for open-weight LLMs

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
large language models, LLM safety, refusal ablation, adversarial training, adaptive attacks, open-weight models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ZMZXOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether safety-trained open-weight LLMs remain robust when attackers jointly remove multiple refusal directions rather than a single one. The authors first propose Refusal Subspace Ablation (RSA), an efficient adaptive attack that iteratively selects complementary directions evaluated under joint ablation, exposing vulnerabilities missed by single-direction tests. Building on this, they introduce Refusal Subspace Ablation Training (ReSAT), which combines mixed supervised fine-tuning with iterative safety-training rounds that supervise safe refusals under cumulative subspace removal, refreshing candidate directions between rounds. Evaluations on three model families show that ReSAT achieves lower attack success rates than most baselines under various ablation and jailbreak attacks while largely preserving general capability, with mechanism analysis confirming the value of retaining direction history.

Critical strengths

S1. Clear presentation throughout. The paper is easy to follow from motivation through method design to experiments, with each section logically building on the previous one. A minor suggestion: the visual polish of tables and figures could be improved.

S2. A relatively complete research. The paper starts from a concrete gap in existing work (single-direction robustness does not imply subspace robustness), proposes an attack (RSA) that exposes this gap, then develops a defense (ReSAT) and validates it with thorough experiments and mechanism analysis.

Critical weaknesses

W1. The safety improvement may stem from increased sensitivity rather than genuine robustness. On XSTest (Table 2), all three ReSAT-trained models exhibit over-refusal rates near 50%, indicating substantial exaggerated safety behavior on benign prompts. This suggests the defense may work partly by making the model more broadly cautious rather than by genuinely preserving refusal under adversarial ablation.

W2. Evaluation is limited to non-thinking models. All three evaluated backbones (Llama-3-8B, Mistral-7B, Qwen2.5-14B) are standard instruction-tuned models. Given that reasoning/thinking models have become mainstream, it remains unclear whether ReSAT's defense transfers to settings where refusal behavior is mediated through long internal reasoning traces rather than direct first-token responses.

W3. Defense underperforms baselines in some settings. For example, on Mistral under SOM attack, ReSAT achieves 28.57% ASR compared to MixAT's 4.40%.

Questions for authors

See weakness1 & 2

3 comments

W1 Response. The trade-off should also be compared with safety-training baselines.** Thank you for providing this comment. Balancing safety and helpfulness is a long-standing challenge in alignment, as discussed by Bai et al. (2022) and the XSTest study of exaggerated safety by Röttger et al. (2024). Huang et al. (2025) further document a “safety tax” in Safety Tax: Safety Alignment Makes Your Large Reasoning Models Less Reasonable: in their studied sequential training pipeline, safety alignment restores safety but degrades reasoning capability, which is a longlasting problems and some other baselines are even worse than us for over-refusal and worse than our safety as well.

Thank you for providing this comment. Table 2 in the paper reports XSTest benign compliance (higher is better). On Llama-3-8B, ReSAT achieves the highest benign compliance among the compared safety-training baselines: 53%, compared with 48% for DeepRefusal, 49% for MixAT, and 26% for CAT. At the same time, ReSAT achieves substantially lower RSA endpoint HarmBench ASR: 11.20% versus 23.44%, 82.03%, and 45.31%, respectively (submitted Tables 1–2). Thus, on Llama-3-8B, ReSAT improves refusal-ablation robustness without exhibiting worse over-refusal than the compared baselines. On Mistral-7B, the trade-off also remains competitive: ReSAT retains 57% XSTest compliance, compared with MixAT’s 62%, while reducing RSA endpoint ASR from 65.62% to 20.00%. ReSAT additionally achieves higher MMLU (59.22% versus 52.93%) and GSM8K (42.99% versus 34.27%). These results demonstrate a favorable safety–utility trade-off rather than uniformly worse utility or safety relative to existing baselines.

W2 Response. We thank the reviewer for making this commnet. Our original comparison evaluated ReSAT and the defense baselines on standard instruction-tuned backbones. To test whether the defense transfers to reasoning models, we also trained and evaluated ReSAT on Qwen3-8B with native thinking enabled. Prior defense studies used a setting similar to our original comparison. Bui et al. (AGOPNs, Findings of IJCNLP-AACL 2026) evaluated their defense on three Llama Instruct checkpoints, and Shen et al. (Safety Instincts, ICLR 2026) evaluated SIRL on Llama and Qwen2.5 Instruct models. Xie et al. (Findings of EMNLP 2025) also evaluated DeepRefusal, one of our baselines, on instruction-tuned models.

Using our RSA to attack thinking mode, adaptive RSA raised HarmBench ASR from 9.6% to 88.8% (Table 8). Lexical refusal rate fell from 68.8% to 14.4%, and MMLU accuracy went from 79.75% to 78.25%.

Table 8: RSA on Qwen3-8B with native thinking enabled, before ReSAT training.

ConditionHarmBench ASR ↑Lexical RR ↑MMLU accuracy
Clean thinking9.6%68.8%79.75%
RSA )88.8%14.4%78.25%

Using our ReSAT safe training for thinking mode. With native thinking enabled, HarmBench ASR was 11.20% without ablation and 8.00% under both Single and cumulative RSA (Table 9). The ReSAT trained thinking model resisted both refusal ablations in this evaluation. We also observed XSTest compliance was 100.00%, no overrefusal on the benign prompts evaluated.

Table 9: ReSAT on Qwen3-8B with native thinking reasoning enabled. Clean, Single, RSA (ours), and HumanJB report HarmBench ASR. XSTest reports benign compliance and MMLU reports accuracy.

ModelClean ↓Single ↓RSA ↓XSTest ↑MMLU ↑HumanJB ↓
ReSAT Qwen3-8B (thinking)11.20%8.00%8.00%100.00%81.82%14.81%

W3 Response. We agree that MixAT provides stronger protection against SOM on Mistral-7B. Its HarmBench ASR under SOM is 4.40%, compared with 28.57% for ReSAT. Under RSA, the ordering reverses. MixAT reaches 65.62% ASR, compared with 20.00% for ReSAT. Also, for model in deployment, an attacker always choose the method that works best against that model. We therefore also compare the highest ASR observed for each defense across the four refusal-ablation attacks in Table 1 of the submission (Single, Geometry, SOM, and RSA). MixAT's highest ASR is 65.62% under RSA. ReSAT's highest is 28.57% under SOM, a difference of 37.05 percentage points in ReSAT's favor.

Thanks to the authors for the reply, which addressed some of my concerns, especially weakness 2. Overall I'm inclined to accept this paper; there are still a few minor issues, but I'll keep my current score.

Thank you very much for taking the time to read our rebuttal and for continuing to engage in the discussion. We are pleased that we have addressed all of your major concerns. If there are any remaining questions or concerns, we would be happy to address them further.

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

The paper studies refusal-ablation attacks on open-weight LLMs. It argues that robustness to removing a single refusal direction does not imply robustness to jointly removing several complementary directions. It proposes (1) RSA, a cheap greedy attack that iteratively adds directions orthogonal to the already-selected subspace, scoring each candidate by the effect of the full joint ablation under a harmless-prompt KL constraint; and (2) ReSAT, a LoRA-based defense. ReSAT starts with mixed SFT (refusals, continuous/PAP input attacks, utility and hard-benign data), then runs SafeTrain rounds. Each round re-runs RSA on the current checkpoint, appends a new complementary direction to an accumulated bank, and trains safe refusals while the whole bank is ablated.

Critical strengths
  1. The attack evaluation shows that single-direction robustness does not imply joint-ablation robustness. Several safety-trained models that resist single-direction ablation (e.g., Mistral MixAT rises from 6.25% to 65.62% ASR under rank-three RSA) fail under joint ablation. This matters because it changes how the community should evaluate refusal-ablation defenses. The attack is also cheap: RSA takes about 10 minutes on Llama versus roughly 76 and 111 minutes for Geometry and SOM, which makes it practical as a routine evaluation tool. For a paper whose main value is evaluation, this is the consideration that most supports acceptance.

  2. The mechanism analysis supports the training design. The paper does more than report that accumulation works better than replacement (11.20% vs. 29.07% ASR). It shows that accumulated directions retain 76.0% and 99.9% of their energy after projection onto the historical span, while replaced directions retain only about 0.2%, and that the new directions have positive marginal causal effects. This separates a real claim, that accumulation finds genuinely new complementary directions, from a weaker one, that the method simply works. Results are also replicated across three seeds per model family. This is the strongest evidence that the core idea is not just a tuning artifact, and it partly offsets the incremental novelty concern.

Critical weaknesses
  1. Limited novelty. It combines multi-direction ablation (Geometry, SOM) with refreshed-direction training (ReFAT). The only new element is requiring orthogonal directions and ablating the whole bank during training.

  2. Thin utility evaluation. Only MMLU and GSM8K are used, which are short-answer tasks that miss long-form and instruction-following degradation. XSTest over-refusal rises sharply, from about 95–99% to 47–57%, yet this is treated as minor.

  3. Jailbreak regression. On Qwen, GCG rises from 12.89 (base) to 52.48, and PAIR also gets worse. This holds across three seeds and is not analyzed, so the claimed complementarity to jailbreak defenses does not hold for Qwen.

Questions for authors

See the weakness

4 comments

W1: Limited novelty relative to Geometry, SOM, and ReFAT.

W1.1 Attack (RSA). We thank the reviewer for asking how our approach differs from prior work. Geometry learns refusal directions and concept cones through gradient-based optimization, while SOM searches over ordered combinations of directions derived from a self-organizing map. RSA asks whether an additional direction exposes a vulnerability beyond those already identified. It evaluates each candidate jointly with the retained directions, subject to a KL constraint on benign outputs (Section 3 in paper). On Mistral-7B trained with MixAT, RSA achieves a HarmBench ASR of 65.62%, compared with 0.00% for Geometry and 4.40% for SOM (Table 1 in submission paper), higher than exisiting baseline. In the timing comparison on the Llama-3-8B ReSAT, RSA search took 10.30 minutes, compared with 76.30 minutes for Geometry and 111.33 minutes for SOM (Figure 2 in paper), which is the most efficient one

W1.2 Defense (ReSAT). ReFAT recomputes layerwise refusal features as the model changes during training (Yu et al., 2025). ReSAT also refreshes directions on the updated checkpoint, but retains previously discovered directions and trains against their joint removal (Section 4.1). To isolate the contribution of retaining this history, we compared two Llama-3-8B training runs with the same initialization, data quotas, update budget, and refresh schedule. Replacing the previous directions at each refresh yielded a rank-four HarmBench ASR of 29.07%. Retaining them reduced ASR to 11.20%, a 17.87 percentage-point reduction (Figure 3). A separate ablation found that repeatedly training on the first direction yielded 31.20% ASR at the same evaluation rank (Table 3 in submission paper).

W2: Limited long-form utility and substantial XSTest

W2.1 Long Form Utility Evaluation We thank the reviewer for giving this comment. We evaluate 3 utility benchmarks following the evaluation convention used by other prior safety-training baselines, our main utility reports MMLU, GSM8K, and XSTest; our original evaluation therefore follows the community field's standard protocol. To address the reviewer's interest and request on long form utility, we nevertheless evaluate and report all results among four main compared baseline methods on AlpacaEval2 and LongBench-Write as follow:

Long-form utility and robustness to RSA. AlpacaEval2 reports win rate (%), and LongBench-Write scores response quality on a 1–5 scale. HarmBench ASR is reported at the last accepted RSA rank (Table 1 of the submission).

ModelAlpacaEval2 (%) ↑LongBench-Write ↑RSA endpoint HarmBench ASR (%) ↓
Llama-3-8B
Base15.592.966792.19
ReFAT14.292.822292.19
MixAT11.242.711182.03
ReSAT (Ours)11.062.616711.20
Mistral-7B
Base11.683.100069.53
ReFAT12.112.838928.91
MixAT8.942.638965.62
ReSAT (Ours)9.502.616720.00
Qwen2.5-14B
Base30.253.322279.69
ReFAT16.522.838964.06
MixAT17.452.661129.69
ReSAT (Ours)17.642.86678.80

Utility performance depends strongly on the utility data included during training. ReSAT's utility branch used the short-form mixture specified in the paper and was not optimized on long-form utility data.If long-form quality is the target, it can be further improved by adding matched long-form utility examples during SafeTrain, which is orthogonal to ReSAT's refusal-subspace robustness mechanism. The present new long form utility results could be read as transfer and generalize to an unseen utility test.

W2.2 The trade-off should also be compared with safety-training baselines. Thank you for raising this concern. Balancing safety and helpfulness is a long-standing challenge in alignment, as discussed by Bai et al. (2022) and the XSTest study of exaggerated safety by Röttger et al. (2024). Huang et al. (2025) further document a “safety tax” in Safety Tax: Safety Alignment Makes Your Large Reasoning Models Less Reasonable: in their studied sequential training pipeline, safety alignment restores safety but degrades reasoning capability, which is a longlasting problems and some other baselines are even worse than us for over-refusal and worse than our safety as well.

On Llama-3-8B, ReSAT achieves 53% XSTest compliance, compared with 48% for DeepRefusal, 49% for MixAT, and 26% for CAT. Its last accepted RSA endpoint ASR is also lower: 11.20%, versus 23.44%/82.03%/45.31%, respectively (submitted Tables 1–2, pp. 6 and 8). On Qwen2.5-14B, ReSAT achieves higher XSTest benign compliance than CAT and MixAT (47% versus 44% and 41%, respectively), while providing substantially stronger protection than the Base model against both single- and multiple-direction refusal ablation attacks (Tables 1–2 in submission paper).

W3 Reviewer concern : Jailbreak regression on Qwen: GCG and PAIR worsen across three seeds, challenging the claimed complementarity with jailbreak defenses.

W3: Jailbreak regression and cross-backbone evidence Response. Thank you for providing this comment. Qwen exhibits a different empirical pattern from the other evaluated backbones, but ReSAT still substantially improves its refusal-ablation defense: RSA HarmBench ASR drops from 79.69% to 8.80%, compared with 72.66% for DeepRefusal (submitted Table 1). Other baselines like DeepRefusal also has even worse Qwen's GCG/PAIR scores from 12.89/31.01 to 32.58/43.18 and deeprefusal has much weaker refusal-ablation defense than ReSAT(submitted Table 2). In contrast, ReSAT improves GCG/PAIR results on Llama and Mistral generally. To address your concern, we also run a Gemma test and the resuls are as followed(Table 7 below). It performs well on Gemma, Llama, and Mistral, whereas Qwen exhibits a bit distinct empirical pattern, potentially reflecting model-family-specific differences.

Gemma-7B-it: refusal-ablation attacks, jailbreaks, and utility. ASR is measured by HarmBench

Model / Gemma-7BClean (Attack ASR ↓ / MMLU ↑)Single (Attack ASR ↓ / MMLU ↑)RSA (Attack ASR ↓ / MMLU ↑)HumanJB (Jailbreak ASR ↓)GCG (Jailbreak ASR ↓)MMLU-400 ↑XSTest ↑
Base3.20 / 46.7588.80 / 47.2588.80 / 47.256.63246.7592
CAT2.40 / 53.7536.00 / 53.2563.20 / 52.005.22854.5019
LAT0.80 / 46.2571.20 / 47.7572.00 / 47.250.82246.5064
DeepRefusal0.00 / 44.0046.40 / 43.7546.40 / 43.750.81443.2578
ReSAT (Ours)0.00 / 49.259.60 / 50.757.20 / 49.500.0249.5061

Thank you for the additional results. However, ReSAT ranks lowest or near-lowest on AlpacaEval2 and LongBench-Write for Llama and Mistral, so I read this as a utility cost rather than “generalization”. On XSTest, the comparison covers only Llama and part of the baselines on Qwen. On Mistral, ReSAT’s 57 is the lowest of all methods. Therefore, I will keep my score.

Thank you for your follow up and considering our additional results. We understand that the only thing left seems just the utlity cost. Our results show that this cost is not uniform across settings: compared with other safety-training baselines, ReSAT maintains or improves some utility metrics. In several comparisons, substantial gains in safety robustness are accompanied by only modest utility reductions, highlighting a meaningful utility–robustness trade-off.

We also want to mention that utility performmance depends strongly on the utility data included during training. ReSAT's utility branch only used the short-form mixture specified in the paper and was not optimized on long-form utility data.If long-form quality is the target, it can be further improved by adding matched long-form utility examples during SafeTrain. If the reviewer see this very important, we would like to run another version involved long form data.

Reviewer_62YUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper tackles the problem of training robust LLMs that can alleviate downstream adversarial attacks. The paper first proposes a new attack method that jointly ablates several refusal directions, and then proposes a defense against this attack. Several experiments have been conducted to showcase its efficacy.

Critical strengths
  1. The research topic of designing robust Open-weight language models is important.
  2. The paper is clearly written and easy to understand.
Critical weaknesses
  1. The method itself is concerning. ReSAT is designed to defend against RSA-based attacks and uses the same probe token, candidate pool, and greedy orthogonal expansion procedure as RSA, so how well the defense generalizes remains unknown.
  2. The results are concerning in terms of stability. The main experiments do not report mean and variance over multiple runs. Note that in Table 14, the Geometry ASR on Qwen2.5-14B is 7.14, 77.92, and 2.60 across seeds. In addition, the method has many hyperparameters but lacks ablation studies on them. Also, utility evaluation is limited to MMLU, GSM8K, and XSTest.
Questions for authors
  1. How does ReSAT perform under RSA with other probes (e.g., Sorry, I+Sorry) or other candidate pools?
  2. Why does Qwen’s Geometry ASR vary so much across seeds (7.14 / 77.92 / 2.60)?
  3. How sensitive is ReSAT to its key hyperparameters (e.g., the loss weights λ and the perturbation radius)?
3 comments

W1 Generalization to other probes and candidate pools?

W1.1 Novelty Response. We thank the reviewer for providing this comment. We have discussed this in the paper Section 5.2. Although ReSAT is trained against RSA, its evaluation is not restricted to RSA. As shown in Table 1, we evaluate ReSAT against both single-direction ablation attack method and multi-directional ablation methods,like Geometry and SOM, in addition to RSA, across three model families. ReSAT substantially improves over the corresponding base models and most existing safety-training baselines across these attacks, with very low ASR and much safer than other baselines, indicating the robustness learned through RSA-based training could be transfered and generalized to other direction-ablation methods.

W1.2 Candidate-pool control. For your concern, we additionally freeze each trained ReSAT checkpoint and re-extract attack directions from a disjoint pool. The new pool contains prompts from PKU-SafeRLHF and Alpaca instructions, while the ReSAT(ours) is still very safe with even lower ASR than original pool.

Original versus new pool RSA (HarmBench ASR in %, lower is better)

BackboneOriginal poolNew pool
Llama-3-8B11.201.60
Mistral-7B19.202.40
Qwen2.5-14B8.8010.80

The independently re-extracted pool does not remove ReSAT's robustness: ASR is 1.60% on Llama, 2.40% on Mistral, and 10.80% on Qwen. We will include these new results in the revised manuscript.

W1.3 Probe control. We have already discussed and analysed the Probe-token sensitivity in Appendix A.1 and Appendix B.2.2 (Table 8). For your concern, we further compare the I, I+Sorry, and Sorry-only probes. Adding Sorry to I does not increase ASR on any backbone, while Sorry-only yields no eligible direction when attacking. These results show that ReSAT could also be generalized across different probe variants. We will include these new results in the revised manuscript.

Probe comparison on the original candidate pool ( ASR in %, lower is better). N/A means unavailabe, that no candidate passed the predeclared eligibility gate.

BackboneII+SorrySorry-only
Llama-3-8B11.2011.20N/A
Mistral-7B19.2019.20N/A
Qwen2.5-14B8.808.80N/A

W2 Stability and sensitivity:

W2.1 We thank the reviewer for highlighting the variation across training seeds. For Qwen2.5-14B, Table 14 of the submission reports HarmBench ASRs of 7.14%, 77.92%, and 2.60% under Geometry across three independently trained ReSAT checkpoints. The mean ASR is 29.22%, with a sample standard deviation of 42.24 percentage points. On the same three checkpoints, the sample SDs are relatively stable: 0 percentage points for Single, 3.90 for SOM, and 0.00 for RSA.

To examine the Qwen Geometry result further, we evaluated 3 additional seeds and obtained ASRs of 5.84%, 2.60%, and 24.03%. Table 3 of this response reports all six runs. We will include the additional results and report the per-seed values, means, and sample SDs in the revised manuscript.

Qwen Geometry control across different seeds.

Seed No.123456
ASR7.1477.922.605.842.6024.03
MMLU-40078.5077.2578.7577.2578.5077.50

W2.2 Several training choices raised by the reviewer are shared with existing safety-training methods. Like prior work, we mix safety examples with general utility data. Table 12 in Appendix F of the submission gives the sampling proportions and identifies the settings held fixed in each comparison. For the MixAT-style input perturbation, we use a default radius of 0.075 times the mean embedding norm. On Llama-3-8B, with the loss weights fixed, halving the radius raised RSA endpoint ASR from 11.20% to 16.00%, while doubling it lowered ASR to 9.60%. MMLU accuracy ranged from 64.50% to 65.00% across the three settings.

Perturbation-radius sensitivity on Llama-3-8B. The training radius is c times the mean embedding norm. All runs use the same loss weights (α=1) and 96 training updates. Metrics are percentages.

RadiuscRSA endpoint ASR ↓Lexical RRMMLU-400 ↑
Half0.037516.0080.0064.75
Default0.07511.2084.8064.50
Double0.159.6086.4065.00

Table 3 in Section 5.5 of the submission reports six ablation groups on Llama-3-8B, including the number of training rounds, the use of cumulative refusal directions, and the role of training-time RSA. For example, with the same budget of 96 updates, training repeatedly on the first direction yielded a HarmBench ASR of 31.20%, compared with 11.20% when training on cumulative directions. Table 13 in Appendix H evaluates sensitivity to the KL threshold used for RSA direction selection. We will add all new results to the revised manuscript.

W2 continued: Utility evaluation is limited to MMLU, GSM8K, and XSTest.

W2.3 Utility evaluation.

We evaluate 3 utility benchmarks following the evaluation convention used by other prior safety-training baselines, our main utility reports MMLU, GSM8K, and XSTest; our original evaluation therefore follows the community field's standard protocol. To address the reviewer's interest and request on other utility benchmarks like long form utility, we nevertheless evaluate and report all three backbones among four main compared baseline methods on AlpacaEval2 and LongBench-Write as follow:

Long-form utility and robustness to RSA. AlpacaEval2 reports win rate (%), and LongBench-Write scores response quality on a 1–5 scale. HarmBench ASR is reported at the last accepted RSA rank (Table 1 of the submission).

ModelAlpacaEval2 (%) ↑LongBench-Write ↑RSA endpoint HarmBench ASR (%) ↓
Llama-3-8B
Base15.592.966792.19
ReFAT14.292.822292.19
MixAT11.242.711182.03
ReSAT (Ours)11.062.616711.20
Mistral-7B
Base11.683.100069.53
ReFAT12.112.838928.91
MixAT8.942.638965.62
ReSAT (Ours)9.502.616720.00
Qwen2.5-14B
Base30.253.322279.69
ReFAT16.522.838964.06
MixAT17.452.661129.69
ReSAT (Ours)17.642.86678.80

Table 5 shows that, across all three backbones, ReSAT scores below the base model on both long-form benchmarks but has the lowest RSA endpoint ASR among the four models compared. On Llama-3-8B, ReSAT's AlpacaEval2 win rate is 0.18 percentage points lower than MixAT's (11.06% vs. 11.24%), while its RSA endpoint ASR is 70.83 percentage points lower (11.20% vs. 82.03%). On Qwen2.5-14B, ReSAT scores higher than MixAT on both AlpacaEval2 (17.64% vs. 17.45%) and LongBench-Write (2.8667 vs. 2.6611), and has a lower RSA endpoint ASR (8.80% vs. 29.69%). We will include these results in the revised manuscript.

Utility performance depends strongly on the utility data included during training. ReSAT's utility branch used the short-form mixture specified in the paper and was not optimized on long-form utility data. If long-form quality is the target, it can be further improved by adding matched long-form utility examples during SafeTrain, which is orthogonal to ReSAT's refusal-subspace robustness mechanism. The present new long form utility results could be read as transfer and generalize to an unseen utility test.

Thank you again for your review. We have provided a detailed response to your comments and concerns above. As the discussion period is approaching its end, we would greatly appreciate it if you could take a moment to review our response and let us know whether it addresses your concerns.

We would also be happy to provide any further clarification if needed. Thank you very much for your time and consideration.

Comments on the paper

2 comments
Area_Chair_wkjrOpenReview ↗

Reminder: Please review the author rebuttal before the deadline

Dear reviewers,

The rebuttal period is about to close. If you haven’t responded yet, please read the authors’ rebuttal and leave a brief comment before the deadline.

If you’ve already responded, please disregard this reminder. Thanks for your time, and apologies for the extra email.

Best regards,

Area Chair

We thank the reviewers for their feedback and consideration of our rebuttal. We thank reviewers highlighted RSA’s practical value for safety evaluation on open weight models, showing that models resistant to single-direction ablation can remain vulnerable to joint ablation, while requiring substantially less search time than other baseline like Geometry and SOM. Our ReSAT improves robustness by accumulating complementary refusal directions, providing strong defenses against diverse refusal-ablation attacks across multiple model families while also maintaining strong protection against jailbreak attacks, compared with other baselines.

We appreciate Reviewer ZMZX support for acceptance and acknowledgment that our responses addressed most of their major concerns.

Reviewer gYHm raise the utility issue, we believe this is a long-standing question. Our performance demonstrates a meaningful trade-off: we achieve greater safety than other baselines while maintaining the same or similar utility. Moreover, performance on a target utility task can be improved by incorporating relevant utility data during training. We originally used short-form utility data during training, but our performance on long-form benchmarks is still nearly the same as that of other baselines, while showing better defense against refusal-ablation attacks.

Reviewer 62YU has not responded to our rebuttal. Their main concerns are actually already addressed in the paper, and we have conducted additional experiments such as candidate pool and probe control. to support those points. Regarding the Qwen-specific issues, we added experiments with six seeds and experiments on another model family, Gemma, to show that our attack, RSA, and our defense, ReSAT, can generalize across different model families. We hope the reviewer will take a look at our response. Many thanks

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.