- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes a framework for manipulating region-level attention during inference to investigate the refusal behaviors and safety vulnerabilities of large language models by cutting off or adjusting attention between different semantic regions. The authors conducted experiments on 12 open-weight models and three security benchmarks, and found that some models, even if they performed robustly at the beginning of the response, may still be bypassed and rejected in subsequent generation.
Critical strengths
- The topic holds research value, offering a new perspective on the safety evaluation of open-weight models.
- The paper demonstrates sound overall logic and a clear structure, with the research problem and key contributions explicitly articulated.
- The experiments encompass multiple model families and safety benchmarks, and the main findings are supported by robust experimental evidence.
Critical weaknesses
- Some experimental setups differ from standard usage practices for the models; the scope of applicability of the conclusions requires further clarification.
- Comparisons with existing methods are insufficient, and the relative advantages of the proposed method remain unclear.
- The analysis of performance differences among the models lacks depth, and the explanation of underlying mechanisms needs refinement.
- Some figures and tables lack visual appeal; there is room for improvement regarding layout and the clarity of annotations.
Questions for authors
- If the full reasoning process of the reasoning-oriented model is retained, do the main conclusions still hold? Are the results robust across different chat templates and generation settings?
- Provide a direct comparison with existing internal intervention methods and demonstrate the advantages of the proposed method under comparable experimental conditions.
- Analyze the differences in sensitivity to attention intervention among various models and provide evidence to support the explanation of the underlying mechanisms.
- Optimize the layout, color scheme, and labeling of certain figures to make the differences between experimental conditions more intuitive.
11 comments
HarmBench, ASR (%):
| Model | Decision cut | Continuation | Refusal ablation | CAA |
|---|---|---|---|---|
| Llama-3.2-3B | 43.5 | 56.0 | 50.0 | 33.5 |
| Llama-3.1-8B | 32.5 | 55.0 | 58.5 | 44.5 |
| Llama-3.3-70B | 20.0 | 27.0 | 54.5 | 37.0 |
| Qwen3-8B | 4.0 | 36.5 | 59.5 | 2.5 |
| Qwen3-14B | 2.5 | 28.5 | 6.0 | 2.5 |
| Qwen3-32B | 1.0 | 25.5 | 0.5 | 0.5 |
| Qwen3-2507-30B | 6.0 | 27.0 | 56.0 | 1.5 |
| R1-Distill-7B | 20.0 | 64.5 | 15.5 | 7.5 |
| R1-Distill-14B | 4.5 | 67.5 | 1.5 | 0.5 |
| R1-Distill-32B | 12.0 | 66.0 | 0.0 | 2.5 |
| gpt-oss-20B | 34.5 | 59.5 | 44.0 | 29.5 |
| gpt-oss-120B | 29.5 | 52.0 | 29.0 | 9.5 |
JailbreakBench, ASR (%):
| Model | Decision cut | Continuation | Refusal ablation | CAA |
|---|---|---|---|---|
| Llama-3.2-3B | 30.0 | 56.0 | 38.0 | 39.0 |
| Llama-3.1-8B | 26.0 | 52.0 | 48.0 | 38.0 |
| Llama-3.3-70B | 13.0 | 15.0 | 50.0 | 25.0 |
| Qwen3-8B | 1.0 | 39.0 | 58.0 | 0.0 |
| Qwen3-14B | 1.0 | 18.0 | 0.0 | 1.0 |
| Qwen3-32B | 0.0 | 15.0 | 0.0 | 0.0 |
| Qwen3-2507-30B | 0.0 | 21.0 | 48.0 | 0.0 |
| R1-Distill-7B | 14.0 | 71.0 | 6.0 | 2.0 |
| R1-Distill-14B | 3.0 | 73.0 | 1.0 | 0.0 |
| R1-Distill-32B | 3.0 | 65.0 | 0.0 | 0.0 |
| gpt-oss-20B | 35.0 | 65.0 | 44.0 | 21.0 |
| gpt-oss-120B | 25.0 | 53.0 | 25.0 | 3.0 |
CAA strength sweep on ClearHarm, ASR (%):
| Model | α=1 | α=2 | α=4 |
|---|---|---|---|
| Llama-3.2-3B | 12.3 | 34.1 | 0.0 |
| Llama-3.1-8B | 7.3 | 33.0 | 3.4 |
| Llama-3.3-70B | 11.2 | 28.5 | 19.0 |
| Qwen3-8B | 0.6 | 0.6 | 1.1 |
| Qwen3-14B | 1.1 | 0.6 | 1.1 |
| Qwen3-32B | 1.1 | 1.7 | 1.1 |
| Qwen3-2507-30B | 2.2 | 1.7 | 1.7 |
| R1-Distill-7B | 5.6 | 6.1 | 5.0 |
| R1-Distill-14B | 0.6 | 0.6 | 1.7 |
| R1-Distill-32B | 0.0 | 0.0 | 0.0 |
| gpt-oss-20B | 2.8 | 7.3 | 48.0 |
| gpt-oss-120B | 2.8 | 3.4 | 2.2 |
Refusal ablation remains stronger on Llama-3.3-70B across all three datasets and on some Qwen3 models on HarmBench and JailbreakBench. The comparison therefore shows different patterns of susceptibility across model families, alongside the broader effectiveness of continuation manipulation.
These comparisons establish relative effectiveness and attack requirements. The value of our work also lies in identifying which attention connections contribute to refusal, when they matter, and how these dependencies differ across models. These controlled tests extend prompt-based safety evaluation to internal attention. Other internal attacks examine refusal directions, components, or output distributions. We believe a higher ASR from another method does not invalidate the attention-level vulnerability we identify; different weaknesses can require different defenses.
3. Model differences and underlying mechanisms.
Our explanation is that models differ in how refusal is established at the assistant boundary and response opening, and how it is maintained during continuation. The existing ablations and new measurements provide evidence for these distinct stages.
Decision-region processing.
Appendix F.2 separates attention from the assistant boundary and response opening to the harmful request. ClearHarm mean ASR gains across the 12 models are reported in percentage points (mean ± SEM):
| Region from which attention is cut | Mean ASR gain |
|---|---|
| Complete decision region | 18.25 ± 5.71 |
| Assistant boundary | 12.75 ± 5.95 |
| Response opening | 4.56 ± 3.24 |
The assistant-boundary effect places substantial sensitivity before the first token is sampled. Refusal is therefore affected by prompt processing as well as the computation at the response opening.
We additionally measured baseline attention to the harmful request and the projection of the decision-position hidden state onto the refusal direction. Relative projection drop is defined as
where values above 1 indicate that a positive baseline projection crosses zero. ASR gains are reported in percentage points relative to the matched baseline.
| Model | Baseline attention to request | Relative refusal-signal drop | Decision-link ASR gain (pp) | Continuation ASR gain (pp) |
|---|---|---|---|---|
| gpt-oss-20B | 0.089 | 0.776 | 54.7 | 79.9 |
| Llama-3.2-3B | 0.094 | 0.765 | 43.6 | 50.8 |
| Llama-3.1-8B | 0.100 | 0.934 | 41.3 | 3.9 |
| R1-Distill-7B | 0.352 | 0.837 | 26.3 | 70.4 |
| R1-Distill-32B | 0.132 | 2.843 | 9.5 | 69.3 |
| Qwen3-2507-30B | 0.130 | 1.031 | 8.4 | 34.6 |
| Qwen3-14B | 0.136 | 0.652 | 7.3 | 34.1 |
| R1-Distill-14B | 0.397 | 0.858 | 3.4 | 74.9 |
| Qwen3-32B | 0.104 | 13.606 | 3.4 | 15.1 |
| Qwen3-8B | 0.144 | 1.604 | 2.2 | 38.0 |
The decision cut lowers the measured refusal signal in every model in this analysis, including models with little behavioral change. Reducing that projection alone therefore does not explain the differences in ASR.
Associations with decision-link ASR gain across these models are:
| Measurement | Pearson r (p) | Spearman ρ (p) |
|---|---|---|
| Baseline decision-region attention to the request | −0.298 (0.40) | −0.675 (0.032) |
| Relative refusal-projection drop | −0.350 (0.32) | −0.486 (0.15) |
Models that allocate more baseline decision-region attention to the request tend to have smaller decision-link gains in the rank comparison. The projection-drop magnitude does not show a statistically significant association with gain. These measurements distinguish changing a refusal-related representation from changing the model's generated behavior.
We thank the reviewer for the constructive feedback and for recognizing the research value of our work and the strength of our experimental evidence. We address each concern below.
Appendix D.2 describes the original generation modes: decision-link and continuation experiments skip reasoning for Qwen3 and R1-Distill, while gpt-oss uses an empty analysis channel. Prefill experiments for Qwen3 and R1-Distill instead continue within the reasoning stage.
We have now repeated the decision-cut and continuation experiments with full reasoning enabled and evaluated refusal ablation [1] under the same reasoning modes. ClearHarm ASR is reported in percent, with each entry showing reasoning off → on:
Reasoning strongly suppresses both attacks on gpt-oss and lowers continuation ASR across all tested models. Decision-cut ASR increases on Qwen3 and the larger R1-Distill models, while R1-Distill-7B retains substantial ASR; Reasoning also weakens refusal ablation baseline. We will report these differences explicitly when stating the scope of our conclusions.
We also tested greedy decoding and temperature 1 alongside each model's stored defaults. Decision-cut ASR gains relative to the matched baseline are reported in percentage points:
Llama and R1-Distill retain positive gains across these settings; gpt-oss is much less sensitive under greedy decoding, and Qwen3 remains largely unaffected.
We use each model's native chat template and identify semantic regions from its rendered template (Appendix C), covering four model families. We have not systematically substituted different templates within the same model. The decoding comparisons above test generation settings while retaining the native template.
We have completed comparisons with refusal-direction ablation (Arditi et al., 2024) [1] and contrastive activation addition (CAA; Rimsky et al., 2024) [2] across 12 models and three datasets, using matched models, harmful requests, decoding settings, and evaluators. We tuned the parameters for refusal ablation and selected CAA's best result across strengths 1, 2, and 4. We also tested strengths 6 and 8, where ASR decreased because the responses became nonsensical.
Both baselines require activation collection and a model-specific direction. Our decision cut requires no activation collection, offline calibration, gradients, model-weight changes, fine-tuning, or adversarial prompt optimization. It changes only the specified attention connections while keeping the model parameters and evaluated prompt unchanged.
ClearHarm, ASR (%):