acceptodds
Under review as a conference paper at ICLR 2027

Harm Recognition to Refusal Gap: Why Small LLMs are unsafe?

Abstract

The growing deployment of small language models makes their vulnerability to jailbreaks an important safety concern. A common belief is that small models are less robust to jailbreaks because they have weaker representation, reasoning, and instruction-following abilities, and thus fail to recognize disguised harmful intent. However, we find that small models can still clearly separate harmful from benign inputs in their internal representations. Their failures arise because this harmfulness-related signal does not reliably translate into refusal behavior. Using existing harmfulness and refusal directions, we show that prompts with strong harmfulness-related activations can still elicit compliant or harmful responses. This suggests a bottleneck in the internal conversion from harmfulness recognition to refusal execution, which we call the recognition-to-refusal gap. Across model scales, safety differences are more closely associated with the coupling between these two signals than with refusal strength alone. Motivated by this finding, we propose Coupling-Aware Direct Preference Optimization (CA-DPO). CA-DPO focuses training on jailbreak cases where harmfulness recognition does not lead to refusal, while retaining benign examples to preserve helpfulness and limit over-refusal. Experimental results show that CA-DPO improves model safety while mitigating over-refusal, suggesting that optimizing the recognition-to-refusal pathway can improve safety while preserving useful behavior in safety-critical deployment.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI safety, Small language model, jailbreak defense, over-refsual

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

This paper studies safety issues related to jailbreaking in small language models (SLMs). The authors first study the mechanism underlying jailbreaking in SLMs and present their findings: small models can still clearly separate harmful from benign inputs in model representations. The failure to refuse jailbreaking requests comes from the lack of refusal behavior after identifying harmful inputs. They name this issue the recognition-to-refusal gap, as the bottleneck comes from the internal conversion from harmfulness recognition to refusal actions. They validate their hypothesis across model families and model sizes within the scope of SLMs. They also propose Coupling-Aware Direct Preference Optimization (CA-DPO), which focuses on adversarial data points for which the model can recognize harmful inputs at the latent level but fails to perform refusal actions. Their experiments demonstrate that the method can improve safety while preserving useful behavior on the tested datasets.

Critical strengths

(1) Important and interesting problem: The paper investigates LLM safety issues and the underlying mechanisms in small language models. Understanding both these threats and methods to reduce the risk is important for developing trustworthy LLMs.

(2) Experiments on diverse model families: The experiments cover models from the Qwen3.5, Llama3.2, and Gemma3 series. The broad coverage of model families provides evidence for good generalization of the proposed method.

Critical weaknesses

(1) Potentially biased evaluation of refusal answers. The main evaluation relies on refusal keywords listed in Figure 8 instead of using semantic checks or LLMs as judges for more precise evaluation. It is possible for LLMs to generate these words while still violating the safety requirements.

(2) Lack of recent baselines: In the experiments, the most recent baseline is Scans, which was proposed in 2025. The rest of the baselines were proposed in 2024 or earlier. The paper itself also discusses some recent methods proposed in 2026, such as HARC [1]. More discussion comparing these methods with recent approaches would improve the soundness of this paper.

Reference:

[1] HARC: Coupling Harmfulness and Refusal Directions for Robust Safety Alignment. Preprint, 2026.

Questions for authors

(1) Based on your experiments on larger open-source models such as Qwen3.5 4B, 9B, and 27B, do you observe that the recognition-to-refusal gap consistently correlates with refusal behaviors as model size increases?

(2) In Figure 2, the jailbreak points are more tightly grouped in a smaller area than the clean samples. I am thus curious to see whether there are any other features that can contribute to this difference, for example, input format and length. Could you provide some examples of the formats and a length analysis (mean and standard deviation of token length) for the different input sources?

2 comments

The reviewer is right that a response can include refusal wording yet still fulfill a harmful request. We have therefore added SORRY-Bench semantic evaluation of the jailbreak outputs: a response that fulfills the harmful request is counted as fulfillment even if it contains refusal words, and an unfulfilled harmful request is treated as a refusal. We retain the original keyword-based false-refusal metric for clean and over-refusal inputs, which are non-harmful requests, and report the metrics separately. The new semantic results support the safety improvement of CA-DPO over Base. We also inspected a subset of answers manually and found that keyword-based judgments were generally close to the observed behavior, with few apparent misclassifications.

ModelMethodCipher ↑H4rm3l ↑Hand ↑
Qwen3.5-0.8BBase0.8327.8317.50
Qwen3.5-0.8BSafety SFT15.6773.5033.67
Qwen3.5-0.8BVanillaDPO44.1791.3339.33
Qwen3.5-0.8BSafeDecoding81.3385.6737.17
Qwen3.5-0.8BSCANS86.0081.1776.17
Qwen3.5-0.8BCA-DPO91.1794.1744.67
Qwen3.5-2BBase9.5029.8369.83
Qwen3.5-2BSafety SFT28.1790.8365.33
Qwen3.5-2BVanillaDPO19.8375.0038.00
Qwen3.5-2BSafeDecoding85.6791.3375.83
Qwen3.5-2BSCANS25.6792.3377.50
Qwen3.5-2BCA-DPO95.6795.1777.00
Llama3.2-1BBase20.3333.5048.33
Llama3.2-1BSafety SFT80.1785.1770.67
Llama3.2-1BVanillaDPO77.1779.0067.33
Llama3.2-1BSafeDecoding88.3382.5080.17
Llama3.2-1BSCANS79.3385.1788.83
Llama3.2-1BCA-DPO89.5089.6775.17
Llama3.2-3BBase84.5051.3350.33
Llama3.2-3BSafety SFT86.1762.5077.17
Llama3.2-3BVanillaDPO87.1772.3348.67
Llama3.2-3BSafeDecoding87.0087.3347.83
Llama3.2-3BSCANS83.3388.8356.50
Llama3.2-3BCA-DPO94.1778.6791.00
Gemma3-270MBase2.174.337.67
Gemma3-270MSafety SFT13.1712.8327.67
Gemma3-270MVanillaDPO15.6736.6728.00
Gemma3-270MSafeDecoding14.3342.1742.83
Gemma3-270MSCANS21.6734.1740.83
Gemma3-270MCA-DPO66.6774.6736.50
Gemma3-1BBase0.1724.5019.50
Gemma3-1BSafety SFT4.5011.5043.00
Gemma3-1BVanillaDPO7.1737.6750.50
Gemma3-1BSafeDecoding13.3341.0050.83
Gemma3-1BSCANS0.3317.1739.50
Gemma3-1BCA-DPO27.3358.5058.33

Each entry is Base → CA-DPO semantic refusal rate, in percent. We will present these results alongside the original keyword-based evaluation and output collapse rates in the revision.

2. Missing comparison with HARC

We have added a direct HARC comparison on the same small models and test prompts, assessing semantic jailbreak refusal, false refusal on non-harmful inputs, and output collapse. CA-DPO outperforms HARC on each evaluated jailbreak set while also reducing false refusals and collapse. We will document implementation and training-setting differences in the revision.

ModelMethodClean ↓OKTest ↓OR ↓FR ↓Cipher ↑H4rm3l ↑Hand ↑Collapse ↓
Qwen3.5-0.8BBase0.119.8335.5015.500.8327.8317.503.07
Qwen3.5-0.8BHARC31.5059.8399.8384.1790.8365.6744.5014.83
Qwen3.5-0.8BCA-DPO0.112.837.830.3391.1794.1744.671.07
Llama3.2-1BBase1.113.3312.3321.5020.3333.5048.336.47
Llama3.2-1BHARC12.1724.6797.3371.8389.3385.1774.839.41
Llama3.2-1BCA-DPO0.891.835.3315.3389.5089.6775.174.13

Here, the jailbreak columns are 100% minus the SORRY-Bench fulfillment rate; the non-harmful columns use keyword-based false-refusal rates. All values are percentages.

Follow-up 1. Scale trends in larger models

Appendix Figures 11–13 report the layer-wise harmfulness-related signal, refusal-related signal, and direction cosine for Qwen3.5-4B, 9B, and 27B, respectively. Across the larger scale range, the 27B model exhibits stronger directional coupling than the smaller models, consistent with the scale-related trend discussed in the paper. Adjacent model sizes are not strictly monotonic—for example, 4B is not uniformly more coupled than 2B—so we describe this as an overall scale trend rather than an improvement guaranteed at every size increment.

Directional coupling should also be distinguished from the per-example recognition-to-refusal gap and from the observed refusal rate. The appendix figures provide layer-wise signal statistics and do not alone establish monotonicity of either per-example gap or refusal rate across all model sizes. We will make this distinction explicit in the revision.

Follow-up 2. Input formats and token lengths in Figure 2

Figure 2 is a qualitative visualization, and its jailbreak samples are primarily cipher examples. Appendix Table 4 compares the three attack surface forms and shows a degree of consistency in their harmfulness-related directions, including positive leave-one-family-out separation. We will add an appendix visualization using the balanced, all-family configuration used by the analyses after Section 3.2. The subsequent quantitative analyses use balanced attack pools and an independent evaluation split; their data composition is described in Appendix A. The source-level token statistics below make the possible format and length differences explicit rather than treating Figure 2 as a controlled length-matched experiment.

SourceNumber of samplesToken length (mean ± std.)
Cipher jailbreak200312.27 ± 65.71
H4rm3l jailbreak200240.60 ± 106.03
Hand-crafted jailbreak200161.71 ± 224.65
Clean300100.91 ± 151.11
OKTest20015.15 ± 4.06
OR-Bench20020.27 ± 5.12
FalseReject20024.85 ± 5.09
Reviewer_gVBsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that small language models are more jailbreakable because although they can distinguish between harmful and non-harmful prompts in their representation space, they cannot act on this knowledge as well as larger models and are thus easier to jailbreak. The authors analyze the activations of the models to arrive at this finding and then propose a training method that attempts to mitigate this issue.

Critical strengths
  • The paper addresses an important problem in small language model safety by investigating why models fail to refuse harmful requests despite retaining harmfulness-related information in their internal representations.
  • The distinction between harmfulness recognition and refusal execution provides an interesting perspective on jailbreak vulnerabilities and motivates a more targeted approach to safety training
Critical weaknesses
  • The evaluation relies on keyword-based refusal detection instead of properly assessing whether the responses contain harmful content, e.g. with a semantic judge validated against humans. The authors can resolve this by doing a human study and validating an LLM judge agreement with human annotators.
  • The proposed recognition-to-refusal mechanism is not causally validated. Cosine similarity between activation directions does not establish a causal relationship, and the coupling stays weak after training. The authors could run a causal ablation, e.g. with activation patching, to more thoroughly validate this.
  • The authors interpret the separation of harmful and benign prompts in representation space as evidence of harm recognition, but this separation could instead reflect superficial differences in wording, formatting, or dataset characteristics. Please address this in the paper and explain how you arrive at the conclusion that separability truly measures harm recognition.
  • The evaluation of generalization and utility is limited. Performance drops substantially on held-out attack types, and the authors do not evaluate whether general reasoning and instruction-following capabilities are preserved after training. To resolve this, the authors could evaluate additional attack families, stronger adaptive attacks, and genuinely held-out underlying harmful objectives.
Questions for authors
  • How do the authors validate refusal detection beyond keyword matching?
  • What causal evidence supports the proposed recognition-to-refusal mechanism?
  • How do the authors rule out wording, formatting, or dataset artifacts as explanations for prompt separability?
  • How well does the method generalize to unseen attacks and harmful objectives, while preserving general capabilities?

Additionally, could you make the pull figure (Figure 1 on page 2) a bit more easily readable/visually polished? It looks AI-generated and is a bit hard to read. E.g. on the right it says "Preferionce training".

If these questions/concerns are addressed properly, I would be open to increasing the score.

2 comments

2. Direct evidence beyond direction cosine

Our mechanism claim does not rely on direction cosine alone. Section 3.3 shows that small models can have a harmfulness-related signal while exhibiting a relatively weak refusal signal. Section 3.4 shows that over-refusal has a weaker input-side risk signal than jailbreak, yet produces a comparable refusal-side signal; the two failures therefore exhibit different forms of signal imbalance. Section 3.5 shows that, within the same attack family, refused jailbreaks generally have stronger coupling between harmfulness-related and refusal-related directions than non-refused jailbreaks.

The direct intervention experiment below provides complementary causal evidence: increasing vharmv_{\mathrm{harm}} alone has little effect on refusal, while increasing rrefr_{\mathrm{ref}} increases both jailbreak refusal and non-harmful false refusal. The training-before/after signal analysis, hidden-coupling ablation, sample-selection ablation, and final CA-DPO results provide additional evidence from representation, training, and behavior levels. Together, these analyses support the recognition-to-refusal-gap interpretation without claiming that any single plot proves the entire mechanism.

Intervention (Qwen3.5-2B Base)Jailbreak refusal (%)Over-refuse refusal (%)Clean refusal (%)
None2.347.810.00
Increase harmfulness direction (approximately 3.5-fold)1.568.591.56
Increase refusal direction (approximately 2.5-fold)28.1337.5011.72

3. Could the PCA separation be caused by input format or dataset artifacts?

We agree that input format and data source can affect PCA representations. Figure 2 is therefore intended as a qualitative visualization of the representation geometry. Appendix A specifies the attack categories; the jailbreak portion of Figure 2 is cipher-dominant. Appendix Table 4 compares cipher, h4rm3l, and hand-crafted attacks. Their harmfulness-related directions show a degree of consistency, and the leave-one-family-out separation remains positive when one attack family is held out. Thus, the observed signal is not completely tied to one surface format.

For transparency, the source-level token statistics are:

SourceNumber of samplesToken length (mean ± std.)
Cipher jailbreak200312.27 ± 65.71
H4rm3l jailbreak200240.60 ± 106.03
Hand-crafted jailbreak200161.71 ± 224.65
Clean300100.91 ± 151.11
OKTest20015.15 ± 4.06
OR-Bench20020.27 ± 5.12
FalseReject20024.85 ± 5.09

These differences mean that Figure 2 alone cannot rule out format or length effects. In a subsequent revision, we will add PCA visualizations covering all attack forms under the same balanced configuration used by the analyses after Section 3.2. Sections 3.3–3.5 already use a balanced attack pool, and CA-DPO is evaluated on an independent split; the data composition and splits are detailed in Appendix A.

4. Generalization to unseen attacks and preservation of general ability

CA-DPO training uses jailbreak samples from cipher and h4rm3l. The hand-crafted test set is excluded from data construction and contains manually constructed attacks, so it evaluates transfer to an attack form not used for training. This is a generalization test rather than a claim that the underlying harmful goals are entirely unseen.

For general capability, we report false-refusal and collapse rates and compare each method's clean answers with the corresponding Base answers using BERT semantic similarity. Base self-similarity is 1.0 by construction.

ModelSamplesBaseSafety SFTVanilla DPOSafeDecodingSCANSCA-DPO
Qwen3.5-0.8B1501.0000.8080.8150.8170.1680.821
Qwen3.5-2B1501.0000.8270.8600.8390.7130.875
Llama3.2-1B1501.0000.8230.8610.7080.7530.866
Llama3.2-3B1501.0000.8960.9000.8130.8070.910
Gemma3-1B1501.0000.8510.8610.8240.9220.877

CA-DPO is among the closest methods to the Base responses on most tested models, while preserving low false-refusal and collapse rates. This supports preservation of clean-response behavior, although semantic similarity alone is not a complete measure of correctness, reasoning, or instruction following.

5. Figure 1, AI use, and readability

We apologize for the typo “Preferionce training” in Figure 1 and will correct it to “Preference training.” The AI Use Statement already included in the manuscript states that generative AI tools were used for visual refinement of Figure 1. We will also standardize “Recognition-to-response gap” to the terminology used in the paper, “recognition-to-refusal gap,” and further adjust the text, layout, resolution, clarity, and readability of the figure.

1. Keyword matching is insufficient for judging safe refusal

We agree that keyword matching alone cannot determine whether a jailbreak answer is safe: a response may contain refusal phrases and still provide harmful assistance. We therefore re-evaluated the jailbreak answers with SORRY-Bench. We treat an unfulfilled harmful request as a refusal, including cases where the answer contains no particular refusal keyword. The semantic evaluation is designed for harmful requests and directly measures whether the harmful intent was fulfilled.

Clean and over-refusal examples are non-harmful requests. For them, the relevant quantity is whether the model incorrectly refuses, so we retain the original keyword-based false-refusal evaluation and report it separately. SORRY-Bench is not used for these two groups because its fulfillment judgment is intended for harmful-request/jailbreak evaluation and can introduce error when applied to ordinary helpful requests or over-refusal prompts.

The table reports the new semantic refusal rates for all methods. For each harmful jailbreak request, we treat a request that was not fulfilled as refused; hence refusal rate is 100% minus SORRY-Bench fulfillment rate. These values are distinct from the keyword-based refusal rates in Table 2 of the submitted paper.

ModelMethodCipher ↑H4rm3l ↑Hand ↑
Qwen3.5-0.8BBase0.8327.8317.50
Qwen3.5-0.8BSafety SFT15.6773.5033.67
Qwen3.5-0.8BVanillaDPO44.1791.3339.33
Qwen3.5-0.8BSafeDecoding81.3385.6737.17
Qwen3.5-0.8BSCANS86.0081.1776.17
Qwen3.5-0.8BCA-DPO91.1794.1744.67
Qwen3.5-2BBase9.5029.8369.83
Qwen3.5-2BSafety SFT28.1790.8365.33
Qwen3.5-2BVanillaDPO19.8375.0038.00
Qwen3.5-2BSafeDecoding85.6791.3375.83
Qwen3.5-2BSCANS25.6792.3377.50
Qwen3.5-2BCA-DPO95.6795.1777.00
Llama3.2-1BBase20.3333.5048.33
Llama3.2-1BSafety SFT80.1785.1770.67
Llama3.2-1BVanillaDPO77.1779.0067.33
Llama3.2-1BSafeDecoding88.3382.5080.17
Llama3.2-1BSCANS79.3385.1788.83
Llama3.2-1BCA-DPO89.5089.6775.17
Llama3.2-3BBase84.5051.3350.33
Llama3.2-3BSafety SFT86.1762.5077.17
Llama3.2-3BVanillaDPO87.1772.3348.67
Llama3.2-3BSafeDecoding87.0087.3347.83
Llama3.2-3BSCANS83.3388.8356.50
Llama3.2-3BCA-DPO94.1778.6791.00
Gemma3-270MBase2.174.337.67
Gemma3-270MSafety SFT13.1712.8327.67
Gemma3-270MVanillaDPO15.6736.6728.00
Gemma3-270MSafeDecoding14.3342.1742.83
Gemma3-270MSCANS21.6734.1740.83
Gemma3-270MCA-DPO66.6774.6736.50
Gemma3-1BBase0.1724.5019.50
Gemma3-1BSafety SFT4.5011.5043.00
Gemma3-1BVanillaDPO7.1737.6750.50
Gemma3-1BSafeDecoding13.3341.0050.83
Gemma3-1BSCANS0.3317.1739.50
Gemma3-1BCA-DPO27.3358.5058.33

CA-DPO improves semantic refusal over Base on all three attack sets for every evaluated model, although it is not the best method in every individual cell. We also manually inspected a subset of responses. The observed judgments were generally consistent with keyword matching, with only a small number of apparent mismatches. The revision will report semantic jailbreak refusal, keyword-based non-harmful false refusal, and collapse rate side by side.

Reviewer_371kOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper identifies a recognition-to-refusal gap in small LLMs: harmfulness can be internally recognized but does not reliably trigger refusal. It proposes CA-DPO, which targets high-harmfulness/low-refusal samples and strengthens the coupling between harmful recognition and refusal while preserving helpful behavior. Experiments show improved jailbreak refusal with relatively low over-refusal and response collapse across several small-model families.

Critical strengths
  • The paper provides a clear mechanistic perspective on small-model safety by separating harmfulness recognition from refusal execution, and shows that this mismatch is correlated with jailbreak failures.
  • The proposed coupling-based training is well aligned with this diagnosis: CA-DPO explicitly pushes refusal representations to better follow harmfulness signals, rather than simply increasing a global refusal tendency.
  • Empirically, the method appears effective across multiple small-model families, achieving substantially higher jailbreak refusal while keeping clean refusal, over-refusal, and response collapse relatively low.
Critical weaknesses
  • The core observation is not entirely new. Zhao et al. (2026) already show that harmfulness and refusal are separately encoded, while Chua et al. (2026) further study coupling these two directions for safety alignment.
  • CA-DPO relies heavily on the harmfulness score H(x)H(x) as the signal that should drive refusal. However, the paper itself notes that over-refusal can be triggered by surface associations between benign expressions and harmful contexts. This raises a similar concern for H(x)H(x): if it also captures lexical or topical shortcuts, stronger coupling may transfer these spurious associations into refusal behavior.
  • The baseline set is relatively outdated and does not include recent methods that directly target harmfulness–refusal coupling or the jailbreak–over-refusal trade-off. References

Zhao, J., Huang, J., Wu, Z., Bau, D., & Shi, W. (2026). LLMs Encode Harmfulness and Refusal Separately. NeurIPS 2026.

Chua, S. P., Wu, H., Ma, Q., & Wu, F. (2026). HARC: Coupling Harmfulness and Refusal Directions for Robust Safety Alignment. arXiv:2607.00572.

Questions for authors

What happens when the harmfulness signal is noisy or spuriously activated? Since CA-DPO explicitly increases refusal when H(x)H(x) is high, could errors in H(x)H(x) be propagated into the refusal mechanism? An analysis with artificially noisy H(x)H(x), hard benign safety-related prompts, or matched intent-contrast examples would help clarify the robustness of the proposed coupling objective.

1 comment

1. Relation to Zhao et al. and HARC

We thank the reviewer for pointing out this related work. We agree that Zhao et al. have shown that harmfulness and refusal representations can be separated, and that HARC proposes to align these two types of directions. We do not claim the separation observation or the general idea of coupling these directions as our novel contribution.

Our focus is the difference between the safety mechanisms of small and large language models: why can a small model fail to refuse a jailbreak even when it has already formed a harmfulness-related representation? We analyze the resulting signal imbalance across model scales and construct preference data that targets failure cases with high harmfulness-related and low refusal-related signals. HARC constrains the two directions at the prompt and response positions, whereas its reported experiments do not cover models below 7B. CA-DPO is designed for substantially smaller models and combines targeted sample selection, preference optimization, and an output-stability objective.

We additionally compare the two methods using the same small models, test prompts, and decoding settings. Jailbreak refusal is measured as 100% minus the SORRY-Bench fulfillment rate; the non-harmful columns use the original keyword-based false-refusal metric.

ModelMethodClean ↓OKTest ↓OR ↓FR ↓Cipher ↑H4rm3l ↑Hand ↑Collapse ↓
Qwen3.5-0.8BBase0.119.8335.5015.500.8327.8317.503.07
Qwen3.5-0.8BHARC31.5059.8399.8384.1790.8365.6744.5014.83
Qwen3.5-0.8BCA-DPO0.112.837.830.3391.1794.1744.671.07
Llama3.2-1BBase1.113.3312.3321.5020.3333.5048.336.47
Llama3.2-1BHARC12.1724.6797.3371.8389.3385.1774.839.41
Llama3.2-1BCA-DPO0.891.835.3315.3389.5089.6775.174.13

On both evaluated small models, CA-DPO provides the better overall trade-off among jailbreak safety, non-harmful-input usability, and output stability.

2. Could noisy harmfulness signals amplify false refusals?

We agree that blindly transmitting a spurious H(x)H(x), for example one associated with a topic or lexical artifact, could increase false refusals. CA-DPO limits this risk in two ways. First, the coupling loss is applied only to labeled jailbreak training examples, rather than to clean or over-refusal examples. Second, preference pairs for non-harmful inputs and clean anchors provide opposing supervision that preserves useful answers.

Figure 4 shows that over-refusal inputs have a weaker input-side harmfulness signal than jailbreak inputs, while their refusal-side signal is comparable. Thus, the observed over-refusal behavior cannot be explained simply as the same high-harmfulness-to-refusal process as jailbreak refusal. Table 2 further reports the resulting false-refusal behavior after training.

To test the causal role of the two directions, we directly amplified vharmv_{\mathrm{harm}} or rrefr_{\mathrm{ref}} in Qwen3.5-2B. The intervention magnitudes were approximately 3.5×3.5\times and 2.5×2.5\times, respectively. The table reports refusal rates (%) over 128 examples per condition.

ModelInterventionJailbreak refusal (%)Over-refuse refusal (%)Clean refusal (%)
BaseNone2.347.810.00
BaseIncrease vharmv_{\mathrm{harm}}1.568.591.56
BaseIncrease rrefr_{\mathrm{ref}}28.1337.5011.72
CA-DPONone77.347.030.00
CA-DPOIncrease vharmv_{\mathrm{harm}}83.596.250.00
CA-DPOIncrease rrefr_{\mathrm{ref}}99.2252.3421.88

Increasing the harmfulness-related direction alone does not produce a systematic increase in non-harmful refusals, whereas directly increasing the refusal direction raises both jailbreak refusals and false refusals. The Base model shows the same directional distinction. These results support the separation between harmfulness recognition and refusal execution. They do not establish that every individual non-harmful example has a perfectly clean H(x)H(x); our claim is about the observed mechanism and aggregate behavior.

3. Missing recent and directly relevant baselines

We have added HARC as a direct baseline under the same small-model, test-prompt, and decoding conditions. The comparison above evaluates semantic jailbreak refusal, non-harmful false refusal, and output collapse. CA-DPO obtains higher semantic refusal on the evaluated jailbreak sets while maintaining lower false-refusal and collapse rates. We will report the implementation details, training data, and budget differences for both methods in the revision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.