acceptodds
Under review as a conference paper at ICLR 2027

Widening the Gap: Exploiting LLM Quantization via Outlier Injection

Abstract

LLM quantization has become essential for memory-efficient deployment. Recent work has shown that quantization schemes can pose critical security risks: an adversary may release a model that appears benign in full precision but exhibits malicious behavior once quantized by users. However, existing quantization-conditioned attacks have been limited to relatively simple quantization methods, where the attacker can estimate weight regions that remain invariant under the target quantization. Notably, prior attacks have consistently failed to compromise more popular and sophisticated schemes, limiting their practical impact. In this work, we introduce the first quantization-conditioned attack that consistently induces malicious behavior that can be triggered by a broad range of advanced quantization techniques, including AWQ, GPTQ, and GGUF I-quants. Our attack exploits a simple property shared by many modern quantization methods: large outliers can cause other weights to be rounded to zero. Consequently, by injecting outliers into specific weight blocks, an adversary can induce a targeted, predictable weight collapse in the model. This effect can be used to craft seemingly benign full-precision models that exhibit a wide range of malicious behaviors after quantization. Through extensive evaluation across three attack scenarios and LLMs, we show that our attack achieves high success rates against a broad range of quantization methods on which prior attacks fail. Our results demonstrate, for the first time, that the security risks of quantization are not restricted to simpler schemes but are broadly relevant across complex, widely-used quantization methods.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM security, quantization attacks, post-training quantization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_F8GROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a quantization attack based on outlier injection. To address the limitation of prior quantization attacks that are restricted to static quantization, the authors exploit the sensitivity of quantization algorithms to extreme values by injecting large outliers into selected weight blocks. This forces the remaining weights within each block to collapse to zero after quantization. The attack is combined with a dual-objective fine-tuning strategy to preserve model utility in full precision while activating malicious behavior after quantization. Experimental results demonstrate that the proposed method achieves high attack success rates across several widely used optimization-based quantization schemes.

Critical strengths
  1. Clear and accessible presentation: The paper is logically organized and easy to follow. The proposed mechanism, from outlier injection to the multi-stage fine-tuning procedure, is presented coherently, and the algorithms and figures are generally clear.

  2. Consideration of defenses and security implications: Beyond demonstrating high attack success rates, the paper discusses potential defense mechanisms. In particular, it investigates statistical detection of outliers using the Kolmogorov–Smirnov (KS) test and considers potential mitigation strategies, such as rescaling and zeroing outliers.

Critical weaknesses
  1. Limited novelty and overall contribution: One of the paper's main contributions is extending quantization attacks to advanced quantization techniques, such as AWQ and GPTQ. However, recent work, including LLMQuA [1], has already explored and demonstrated the feasibility of attacking more sophisticated LLM quantization schemes. Although the proposed mechanism of exploiting outliers to disrupt group-wise scaling is straightforward, the overall conceptual novelty appears limited.

  2. Questionable stealthiness and practical threat: In realistic deployment scenarios, as awareness of quantization-related security risks increases, model users or auditors may inspect model weights before quantization rather than relying solely on black-box functional evaluation of the full-precision (BF16) model. The proposed method requires injecting a substantial number of extreme-magnitude outliers into the weights, resulting in pronounced anomalies in the full-precision weight distribution, as illustrated by the KS-test analysis. Such anomalies may be readily detected using relatively simple statistical tests, raising concerns about the stealthiness and practical threat posed by the attack.

  3. Limited attack effectiveness under certain experimental settings: The attack exhibits clear limitations for some model–quantization combinations. For example, in the Mistral-7B-Instruct + Over-Refusal setting, under GGUF Q4_0 quantization, the attack success rate (ASR) of the attacked quantized model is nearly identical to that of the full-precision model (4.3% versus 4.3%). This result suggests that the intended quantization-triggered behavior is not effectively realized in this setting.

  4. Substantial utility degradation under certain settings: The attack causes non-negligible degradation in general capabilities for some model–attack combinations. For example, after the attack is applied to Mistral-7B-Instruct in the Over-Refusal setting, its HumanEval score drops from 36.6 for the original model to 20.7. This result highlights a potential trade-off between attack stealthiness and the preservation of general model capabilities.

[1] LLMQuA: Practical Backdoor Injection on Large Language Model Quantization. WWW 2026.

Questions for authors

See weaknesses.

2 comments

W1: Can you discuss novelty over LLMQuA?

Good question. LLMQuA [3] is a malicious code insertion attack into open-source quantization libraries. (See Section 3.2 and Algorithm 1 of their work). They assume that an adversary modifies a popular library (e.g., GPTQ) such that its rounding pattern is skewed towards malicious behavior, and that the user uses the compromised library on a benign model to activate the attack. In contrast, we aim to (maliciously) finetune a model such that the attack activates under unmodified quantization. We believe our threat model is substantially different. We included the discussion in the PDF (see Section 2).

W2: Can you elaborate on the practical threat stemming from this attack?

Great question. As we briefly mentioned in Section 5.3 and explained in C.2, our attack is indeed detectable based on statistical testing, assuming access to the original version of the attacked model. However, we believe this does not undermine the impact of our work for the following reason.

Detection in a practical setup is challenging. Section 5.3 is based on a setup where the adversary did not take detection into account. We show in Figure 6 that it is possible to achieve promising (yet slightly lower) ASR even with a significantly smaller scaling factor (16x instead of 512x). Building on this, we further show that the 16x-scaled switching block is still below the maximum value of the weight parameter of the layer (Figure 7), making it more challenging to detect it trivially, and (ii) it still achieves high ASR in various 4 bit quantizations (Table 13).

  • The real deployment pipeline does not have such a detection scheme.* To our knowledge, the current model-sharing and deployment pipeline does not integrate such detection-based inspection. Our goal is to highlight the potential risk, and to encourage developing a safe deployment pipeline that mitigates the risk of similar attacks. Importantly, existing defenses, such as noise defense (Section 5.4) and some other defenses (added in Section B.10), fail against our attack. In contrast, we find statistical testing-based detection to be successful under the knowledge that this model is attacked via our outlier insertion, but this statistic does not work for existing PGD-based approaches that will not create such large statistical anomalies. This introduces an additional layer of complexity for the defender’s side, providing deeper insights to design a reliable inspection mechanism to defend against this type of attack as a whole.

W3: Investigate low ASR for, e.g., Mistral, over-refusal, GGUF Q4_0.

That is a very interesting point and hard to answer. As the reviewer points out, we show in Table 9 that the outlier injection process is largely successful even for the case where we did not achieve high ASR. While we believe it involves various complex factors, we provide additional analysis to interpret this.

We calculate the cosine similarity of the final-layer activation between the proxy model and the actual quantized model, and compute its correlation with ASR (Mistral, over-refusal). As shown in Figure 7, we observe a positive correlation between proxy-real cosine similarity and ASR (Pearson = 0.76). While not a strong signal, this indicates that the layers other than the outlier insertion phase are accumulated over layers, which leads to a lower ASR for some methods. This indicates that a stronger control on how each layer is quantized (beyond the switching layer) could further strengthen the attack. We leave it as an interesting avenue for future work.

W4: Explain low HumanEval score for Mistral with over-refusal setting.

Certainly. We observe a general instability when finetuning Mistral, as also observed in prior attacks [1]. In our attack, we see the degradation already after kickstart finetuning, before outlier insertion (see the table below).

Importantly, we do not consider the utility degradation to be a fundamental limitation of our approach for two reasons. First, as we show in Figure 3 on Llama, we can control the outlier scale to balance the utility-ASR. In particular, if we allow a lower ASR for 8-bit quantization and focus on 4-bit, we achieve promising ASR even with a 16x smaller scaling factor. Second, we use the KL regularizer throughout training to maintain utility. If the adversary’s priority is to preserve utility even more, increasing the coefficient provides the flexibility to do so.

ModelGSM8KHellaSwagMMLUARC-CHumanEval
Attacked49.374.858.053.336.6
After Kickstart Finetuning42.482.560.056.025.6
Final34.672.056.545.420.7

Thanks to the author's detailed reply and supplementary experimental analysis, these results help to provide a clearer understanding of the impact of threat model settings, scaling factor trade-offs, and fine-tuning on model performance. However, considering that other reviewers have raised similar concerns, especially regarding attack concealment, limited attack effects in some settings, and reduced general capabilities, after a comprehensive weighing, I am inclined to maintain the current score.

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

This paper studies quantization-conditioned attacks in which a released full-precision LLM behaves benignly but becomes malicious after quantization. The method injects large outliers into a selected weight matrix and uses dual-behavior fine-tuning with a sparse proxy to exploit the resulting loss of non-outlier weights. Experiments cover three LLMs, three attack scenarios, and a broad set of quantizers, including GPTQ and AWQ. The paper also evaluates utility, sensitivity to user-side settings, and potential detection and mitigation strategies.

Critical strengths
  1. Outlier insertion provides a simple mechanism for inducing quantization-dependent changes without optimizing through each target quantizer. Table 1 demonstrates that a single attacked checkpoint per model and scenario can trigger malicious behavior across multiple quantization families, with particularly strong results on GPTQ and AWQ. This is a meaningful extension of the quantization-conditioned attack setting.

  2. The pipeline ablations and non-outlier zeroing measurements support the importance of the proposed components. The paper also examines user-side quantization settings and reports both the failure of a prior noise-based defense and the detectability of its own weight anomalies. These analyses help distinguish the demonstrated vulnerability from its practical limitations.

Critical weaknesses
  1. The reported average relative utility masks substantial losses on individual tasks. For example, Mistral over-refusal reduces GSM8K from 49.3 to 34.6 and HumanEval from 36.6 to 20.7 (Table 12). The five benchmarks provide useful evidence, but do not establish broad recovery of the capabilities affected by FFN reinitialization. This limits the evidence that the released model would remain sufficiently useful to pass ordinary behavioral evaluation.

  2. The attack creates a pronounced weight-distribution anomaly: Figure 5 identifies the attacked layer relative to the original model and a benign fine-tuning baseline. Thus, low pre-quantization ASR does not by itself establish stealth under weight inspection. Although the authors acknowledge this limitation, the evaluation does not characterize the attack–detection tradeoff across diverse benign models at controlled false-positive rates, leaving its practical threat under such screening insufficiently established.

  3. The zeroing statistics alone do not explain the variation in attack success across quantizers. Table 9 reports 100% non-outlier zeroing for AutoRound, while Table 1 includes much lower ASR than GPTQ for some model–scenario combinations. The paper does not sufficiently characterize the additional discrepancies between the proxy and actual quantized models that determine these outcomes. This limits the explanatory scope of the mechanism and the evidence for reliable transfer across quantization methods.

Questions for authors
  1. Table 12 shows substantial task-specific utility losses in some settings, such as Mistral over-refusal on GSM8K (49.3 to 34.6) and HumanEval (36.6 to 20.7). How much of this degradation comes from reinitializing and recovering the FFN block, rather than the subsequent attack training? A matched baseline that only reinitializes the block and performs benign recovery, together with evaluation beyond the current five benchmarks, would help clarify how broadly the original capabilities are preserved.
  2. Figure 5 shows that the attacked layer is readily distinguishable from the original model and the benign fine-tuning baseline by its weight distribution. How does this affect the practical threat in deployments that inspect weights before quantization? Can the attack retain high success while avoiding detection at a controlled false-positive rate across a broader set of benign fine-tuned models?
  3. When does the sparse proxy predict attack success? Table 9 reports 100% non-outlier zeroing for AutoRound, yet Table 1 shows much lower jailbreak ASR for AutoRound than for GPTQ on Llama3.1. What determines whether the sparse proxy transfers to a particular quantizer beyond the zeroing rate? Using the same attacked checkpoint, could you compare the proxy and actual quantized models to explain the contrasting outcomes?
2 comments

We thank the reviewer for their thoughtful and constructive feedback. We are pleased that the reviewer recognizes the simplicity and transferability of our attack, and the value of our additional analyses. We address each of the concerns below.

W1 & Q1: Can you explain the utility degradation of Mistral, over-refusal, GSM8K & HumanEval (Table 12).

Certainly. We observe a general instability when finetuning Mistral, as also observed in prior attacks [1]. In our attack, we see the degradation already after kickstart finetuning, before outlier insertion (see the table below).

Importantly, we do not consider the utility degradation to be a fundamental limitation of our approach for two reasons. First, as we show in Figure 3 on Llama, we can control the outlier scale to balance the utility-ASR. In particular, if we allow a lower ASR for 8-bit quantization and focus on 4-bit, we achieve promising ASR even with a 16x smaller scaling factor. Second, we use the KL regularizer throughout training to maintain utility. If the adversary’s priority is to preserve utility even more, increasing the coefficient provides the flexibility to do so.

ModelGSM8KHellaSwagMMLUARC-CHumanEval
Original49.374.858.053.336.6
After Kickstart Finetuning42.482.560.056.025.6
Final34.672.056.545.420.7

Based on the reviewer’s suggestion, we additionally provide an evaluation of the models with the IFEval benchmark to provide more demonstrations of utility preservation. As shown, the utility results are largely consistent with the 5 reported benchmark results in the model (i.e., a slight decrease that is particularly visible for Mistral)

ModelOriginalJailbreakContent InjectionOver-Refusal
Llama-3.1-8B-Instruct74.367.161.463.0
Qwen2.5-7B-Instruct70.464.159.561.6
Mistral-7B-Instruct-v0.349.543.835.134.4

[1] https://arxiv.org/abs/2510.07985

W2 & Q2: Could you discuss its practical detectability?

Great question. As we briefly mentioned in Section 5.3 and explained in C.2, our attack is indeed detectable based on statistical testing, assuming access to the original version of the attacked model. However, we believe this does not undermine the impact of our work for the following reason.

Detection in a practical setup is challenging. Section 5.3 is based on a setup where the adversary did not take detection into account. We show in Figure 6 that it is possible to achieve promising (yet slightly lower) ASR even with a significantly smaller scaling factor (16x instead of 512x). Building on this, we further show that the 16x-scaled switching block is still below the maximum value of the weight parameter of the layer (Figure 7), making it more challenging to detect it trivially, and (ii) it still achieves high ASR in various 4 bit quantizations (Table 13).

  • The real deployment pipeline does not have such a detection scheme.* To our knowledge, the current model-sharing and deployment pipeline does not integrate such detection-based inspection. Our goal is to highlight the potential risk and to encourage developing a safe deployment pipeline that mitigates the risk of similar attacks. Importantly, existing defenses, such as noise defense (Section 5.4) and some other defenses (added in Section B.10), fail against our attack. In contrast, we find statistical testing-based detection to be successful under the knowledge that this model is attacked via our outlier insertion, but this statistic does not work for existing PGD-based approaches that will not create such outliers. This introduces an additional layer of complexity for the defender’s side, providing deeper insights to design a reliable inspection mechanism to defend against this type of attack as a whole.
ASR
Attacked1.7
NF413.7
FP490.0
GPTQ 4-bit15.0
AWQ 4-bit92.3
GPTQ 8-bit1.0
LLM.int81.3

W3 & Q3: Can you analyze why the attack works less on some models-scenario combinations despite having 100% zeroing in Table 9?

That is a very interesting point and hard to answer. As the reviewer points out, we show in Table 9 that the outlier injection process is largely successful even for the case where we did not achieve high ASR. While we believe it involves various complex factors, we provide additional analysis to interpret this.

We calculate the cosine similarity of the final-layer activation between the proxy model and the actual quantized model, and compute its correlation with ASR (Mistral, over-refusal). As shown in Figure 7, we observe a positive correlation between proxy-real cosine similarity and ASR (Pearson = 0.76). While not a strong signal, this indicates that the layers other than the outlier insertion phase are accumulated over layers, which leads to a lower ASR for some methods. This indicates that a stronger control on how each layer is quantized (beyond the switching layer) could further strengthen the attack. We leave it as an interesting avenue for future work.

Reviewer_JU7jOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors proposed an extension of prior work to create backdoor that only activates when quantized. The main contribution is a backdoor scheme that transfer across quantization schemes. The attack was tested on 3 LLMs and the results show high ASR.

Critical strengths

I think this is a good work. The algorithm of finding the right switching block, zero-ing weights ("outlier") is interesting. If the code is released, this could be a nice addition to the backdoor community.

Critical weaknesses

Overall, I think this paper is an improvement on backdoor in quantized model, but it lacks technical depth in the analysis. Importantly, since their previous version (of the same paper), many papers were published to fix quantization-conditioned backdoor (QCB). It is paramount to update this work w.r.t. recent defenses (newest in August 2026) (e.g., https://arxiv.org/abs/2606.29239, https://openreview.net/pdf?id=CEfr3h68KU).

  • Claim to be first to attack PTQ but what about https://arxiv.org/pdf/2312.02220 or https://ieeexplore.ieee.org/abstract/document/10418142/, or https://dl.acm.org/doi/pdf/10.1145/3774904.3792256
  • Select one layer and a block as the switching block: how? give technical detail and sensitivity analysis.
  • Table 1: why the 10% ASR threshold? To me it should be at least 50%, no? Or use heatmap color. Everything green loses information. For example, AutoRound looks more robust, and GGUF IQ4_XS isn't too bad either. Also, Mistral 7B is harder to attack (we need explanation!). But in the current visual, you don't see that (doesn't jump out). Maybe you could also integrate the prior attacks into the table to show the delta.
  • Define "stealthiness" of the attack. How would it be detected and why stealthiness matters?
  • As c (scaling) is important, explain it in Step 3 of Section 4.2.
  • Editorial: Figure 6, Table 8 are in Appendix. You should mention it.
  • Analysis doesn't explain why the attack works better on some models than others. Is there any fundamental reasons for that? The authors only tested on 3 LLMs, how does it generalize to larger models? Even if I appreciate the ablation, it could be moved to Appendix because it doesn't make too much sense. It is obvious that without selecting the switching block, the kickstart or the outlier, there are no backdoor…Though the interesting results is for over refusal where it seems that zero init, kickstart, the ASR improves…why?
Questions for authors

Adding all the explanations asked earlier could affect my assessment.

2 comments

We thank the reviewer for their thorough review. We are happy to hear that they think it is good work with a clear presentation, and are surprised to see a score of 1. We address each point below.

W1: Could the authors show ASR under recent defenses?

That’s an interesting point. Beyond the noise-based defense we investigated in Section 5.4, we additionally evaluated the defense methods. Based on the reviewers' comments, we now added QuantGuard [1] and LAC [2], and report the scores in Table 12 (also pasted below).

As shown, neither defense successfully defends against our attack. Importantly, this result is intuitively expected. For example, QuantGuard explicitly targets PGD-based attacks and aims to reverse rounding decisions near quantization boundaries, assuming they are critical to attack activation. In contrast, our attack introduces large outliers that dominate the quantization scale and cause widespread weight collapse, which works fundamentally differently from prior PGD-based attacks. This suggests the need for defenses against a broader range of quantization-triggered attacks.

ASR after NF4
NF4 w/o defense97.0%
NF4 w/ QuantGuard95.3%
NF4 w/LAC98.0%

[1] https://arxiv.org/abs/2606.29239

[2] https://openreview.net/pdf?id=CEfr3h68KU

W2: Is this the first attack on post-transformation quantization (PTQ)?

No, we never claim to be the first attack on PTQ, and we believe there was a misunderstanding. Importantly, we believe we are both aware that there are existing attacks on PTQs in general, as we discussed them in Section 2, and as the reviewer cites in other points. We instead argue that our work is the first that works across “a broad range of quantization methods on which prior attacks fail” (Abstract, line 21-22), which we believe to hold. In case the reviewer considers certain parts of our work unclear in this aspect, we are happy to provide revisions based on their feedback.

Regarding the papers mentioned by the reviewer, we believe their scope is substantially different from ours:

  • QuantAttack [1] is *an adversarial attack on image input. Their objective is to introduce a large number of activation outliers, causing a DoS attack for schemes using a mixed-precision decomposition that separately handles outliers. The assumption only holds for LLM.int8() among the methods we considered.
  • Quantization Aware Attack [2] is an attack on the vision classifier. It aims to introduce an attack that survives (i.e., is robust to) quantization. In contrast, we aim to introduce an attack that is conditioned on quantization, where we aim to maintain low ASR for the full-precision version.
  • LLMQuA [3] is a malicious code insertion attack into open-source quantization libraries, whereas we aim to insert an attack that activates under unmodified quantization.

[1] https://arxiv.org/pdf/2312.02220

[2] https://ieeexplore.ieee.org/abstract/document/10418142/

[3] https://dl.acm.org/doi/pdf/10.1145/3774904.3792256

W3: How are the switching block/layer selected?

Good question. We discussed our findings briefly in Section 5.5, pointing to a more extensive discussion in Appendix B. We summarize them as follows:

  • (Table 8) We found the up_proj matrix of the FFN block to be the most reliable target for inserting outliers, and thus consistently use it.
  • (Figure 6) We ablated the ID of the switching layer and found it to be largely stable except for the first and last few layers. Here, we selected the best layer based on the validation set to balance ASR and utility.

W4-1: Can you improve the color scheme of Table 1? Why did you choose to highlight the cell with ASR > 10%?

We intended to highlight our main message, i.e., that our attack achieves a clearly higher ASR than the full-precision version. For this, we set the threshold at 10% as (one of the) easily interpretable colorings. Prompted by the reviewer’s suggestion, we attempted a gradation instead of the white-green binary coloring. Please see if you find it better than our original version.

W4-2: Could the authors integrate the prior attack into the table?

Interesting point, but we believe this would be suboptimal, as prior works do not apply to a majority of rows. (as also reported as a limitation in prior work [1-3]).

[1] https://arxiv.org/abs/2405.18137

[2] https://arxiv.org/abs/2601.02680

[3] https://arxiv.org/pdf/2505.23786

W5: What is “stealthiness” in this attack, and why does it matter?

Great question. We consider the following aspects to be important to keep the attack stealthy, each of which we reported in the paper.

  • In full precision, the attacked model has similar utility/ASR compared to the base model (Tables 1 and 2)
  • The attacked, quantized model is not broken in utility. (Table 10)

In our threat model (see Section 4.1), we assume an adversary uploads a full-precision model to Hugging Face, on which the model performs comparably to other models. This is why the first point is important. The adversary aims to introduce a targeted behavior for the quantized model (e.g., advertising a brand name) rather than a complete collapse, where the second perspective matters.

W6: Can you better explain the scaling factor cc in Section 4.2 Step 3?

Absolutely. As the reviewer recognizes, the scaling factor is crucial for the attack: it should be large enough to dominate the quantization scale, but not so large that it severely degrades the utility. We attempted to clarify this point (see new Section 4.2). We would be happy to revise further if there are specific aspects the reviewer thinks are missing.

W7: In Section 5.5, mention that Figure 6 and Table 8 are in the appendix.

Certainly. We have clarified this in the updated Section 5.5 and thank the reviewer for raising this point.

W8-1: Could the attack generalize to larger models?

Yes. Based on the reviewers' comments, we evaluated Qwen-2.5-14B-Instruct under the jailbreak setup and report the result in Table 11 (Section B.9). As shown, we observe a strong attack success rate, indicating that the attack is not limited to the smaller models we used for the main experiments.

W8-2: Can you explain why ablating each step matters (Table 5)?

Great question. We summarize our findings based on Table 5.

  1. Without zero-initialization, we observe low utility, highlighting the need for this step from a utility perspective.
  2. While the reviewer finds the kickstart step to be necessary with no room for ablation, the attack could be conceptually possible without this step in our view, as Step 4 still contributes to introducing different behaviors. We empirically validated the reviewer's intuition by ablating this step.
  3. As the reviewer correctly recognized, the outlier insertion step is the most critical in that the attack must succeed. We included its ablation for completeness to highlight this point.
  4. The refinement step contributes to stabilizing the attack. Without it, the attack either has too large an ASR for the full-precision model or a lower ASR for the quantized model.

We clarified this in the updated revision.

W8-3 Again in Table 5, why does dropping a step sometimes have higher ASR?

That’s a good catch. We note that the finetuning step is always a combination of adversarial loss (to increase ASR of the quantized model) and benign loss (to maintain the utlity/ASR of the full-precision model). The influence of each term is model- and scenario-dependent.

For example, in the mentioned over-refusal setup without kickstart, it does achieve higher quantized ASR, yet at the cost of even higher full-precision ASR. Only with our full pipeline does the attack succeed across all metrics.

Comments on the paper

1 comment

We thank the reviewers for the thorough review! We have made the following updates to the PDF. Changes are marked in blue. We mention relevant changes in the individual responses.

Our changes are as follows

New Experiments. We provide the following new experiments in the appendix.

  • Table 11: We evaluate ASR on a larger model, Qwen2.5-14B-Instruct.
  • Table 12: We evaluate the attack against the defense methods QuantGuard and LAC.
  • Table 13 and Figure 8: We analyze the detectability and ASR of the attack with a small scaling factor.
  • Figure 7: We analyze the relationship between proxy alignment and ASR.

Presentation. We appreciate the reviewers for their suggestions in better presenting the paper. We have adjusted the coloring of the main table.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.