acceptodds
Under review as a conference paper at ICLR 2027

How to Tame a Multi-Headed Hydra? Adaptive Multi-Category Safety Steering for Large Language Models

10 views

Abstract

As large language models (LLMs) become increasingly widespread, preventing unsafe responses to harmful prompts is essential for their safe deployment. Activation steering offers an approach to improving LLM safety by modifying internal activations during inference without updating model parameters. However, a single prompt can involve multiple harm categories, and steering toward safety in one category may leave harmful content from another unaddressed. Despite advances in adaptive steering, existing methods do not explicitly coordinate steering direction and strength when multiple harm categories co-occur within a single prompt. To address this problem, we propose CAM-Steer, a Category-Adaptive Multi-category Safety Steering framework. Specifically, it estimates the risk associated with each harm category by comparing the current hidden state with safe and unsafe prototypes. The estimated risks are then used to combine the safety directions for different harm categories into a single steering direction and to determine the strength of the intervention. Finally, it rotates the hidden state along the composed steering direction, with the rotation angle determined by the estimated risks, while preserving the hidden-state norm. Experiments across three LLM backbones and seven harm categories show that CAM-Steer outperforms the evaluated baselines in average defense success rate, including when categories co-occur. Further analyses support its component designs and informative risk scores, with negligible inference overhead.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM safety, activation steering, multi-category safety, representation engineering, adaptive steering

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_CuzpOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies safety steering when a single prompt can involve more than one harm category. The method, CAM-Steer, estimates a category-wise risk vector from hidden states and uses it both to compose safety directions and to set the strength of a norm-preserving spherical rotation. It is evaluated on BeaverTails’ seven harm categories and three 8–9B open-source models. On mixed-harm prompts, defense success is higher than the authors’ additive and single-direction variants, utility is largely preserved, and ablations suggest that risk modeling and the spherical update each contribute. I read the validation as showing that this combination works on the evaluated benchmark, rather than as establishing a general multi-category steering framework.

Critical strengths
  1. The mixed-harm setting is a concrete operational issue, not only a slogan. A prompt can carry more than one harm, and steering a single category can leave another open. The self-harm-plus-violence example and the multi-harm tables make this visible, which is the main reason I take the problem seriously.

  2. The method is internally consistent, and the evaluation on this benchmark is strong enough to support acceptance of the empirical claim. One risk vector controls both which directions are mixed and how far the representation is rotated. Relative to the authors’ additive and collapsed-direction controls, mixed-harm defense improves, utility stays in a reasonable range, and the ablations move the result. That is the evidence I would rely on if recommending acceptance.

Critical weaknesses
  1. The paper does not isolate a new problem as clearly as the framing suggests. The introduction places safety, activation steering, multi-category leakage, and adaptive coordination of direction and strength side by side, but the links are thin. Related work already includes multi-attribute composition, gating, and rotation, and it is not shown why those methods could not be applied to the same mixed-harm setting. This is consequential because I would accept a working instance more readily than a claim that a new steering paradigm has been identified.

  2. The broader robustness language is ahead of the evidence. The main results stay inside one taxonomy, and the risk scores that coordinate the method are only lightly diagnosed when two harms appear in one prompt. That does not undo the BeaverTails result, but it is why I would not treat the paper as having established a generally reliable multi-category framework.

Questions for authors
  1. Related work already includes multi-attribute composition and rotation-based steering. Is there already a comparison, in the paper or appendix, that places such a method on the same mixed-harm BeaverTails setting? If so, please point to it; that would change how sharply I separate CAM-Steer from an instance of existing ideas.

  2. The risk vector decides both which directions are used and how strong the update is. Do you already report per-category error, or how the scores behave when two harms appear in one prompt? If that analysis is present, I would put more weight on the “coordination” claim. If not, I will keep reading that part as only lightly tested.

  3. Are the main directions and risk estimators built only on BeaverTails’ seven categories, or is there already a result on a different taxonomy or domain? If the latter is already in the paper, the robustness wording would be easier to accept.

4 comments
AuthorsOpenReview ↗

Response to Reviewer Cuzp

W1 & Q1: Positioning against existing steering approaches

A1: We appreciate the reviewer’s comments on the relationship to existing multi-attribute and rotation-based steering methods. We would like to clarify that the submitted paper already includes these approaches in the experimental comparisons. In Section 4.1 (Baselines), we introduce MAT-Steer and K-Steering as multi-attribute control methods, along with ORBIT-R and ORBIT-B as rotation-based baselines. Table 2 reports the results on Qwen, while Appendix Table 10 provides the complete comparison across Qwen, Llama, and Gemma. CAM-Steer outperforms all four methods across the evaluated models.

We also agree that the contribution should be framed more precisely. Rather than claiming a new steering paradigm, our contribution lies in developing and empirically validating a category-adaptive mechanism that coordinates safety directions and intervention strength for co-occurring harm categories. We will clarify this positioning more clearly in the Introduction and Related Work.

W2 & Q2: Multi-label risk-score behavior

A2: We appreciate the suggestion to provide a more direct analysis of the risk scores when multiple harm categories appear in the same prompt. Figure 5 and Appendix Figure 9 examine the category selectivity of the risk scores and their coverage of the source category. We now extend this evaluation to multi-label detection on prompts containing two to five active harm categories. Table R1 reports micro precision, recall, F1, exact match, and all-category recall. The micro F1 reaches 77.22% on Qwen, 70.58% on Llama, and 67.82% on Gemma. The results indicate that the risk estimator captures useful category information while still missing some active categories. However, we note that incomplete category detection does not necessarily lead to unsuccessful defense, as related safety directions may share representational components that help suppress harmful behaviors associated with missed categories. We also report response-level harmful leakage in Table R2.

Table R1. Multi-label detection of the seven modeled harm categories with K=2 to K=5 active categories.

BackboneMicro precisionMicro recallMicro F1Exact matchAll-category recall
Qwen3-8B75.5678.9677.2247.6173.25
Llama-3.1-8B-Instruct67.6773.7570.5834.7166.08
Gemma-2-9B-IT61.0376.3067.8229.7868.95

Table R2. Response-level harmful category leakage across different backbones and steering methods.

BackboneMethodExactly oneExactly twoThree or more
Qwen3-8BCAST122522
Qwen3-8BCAM-Steer7176
Llama-3.1-8B-InstructCAA112729
Llama-3.1-8B-InstructCAM-Steer122120
Gemma-2-9B-ITAdaSteer158
Gemma-2-9B-ITCAM-Steer543

W2 & Q3: Evaluation scope across taxonomies

A3: We thank the reviewer for highlighting the need for a clearer discussion of generalization. The category directions, prototypes, and risk calibration in the main experiments are constructed from the seven modeled BeaverTails categories. The submitted Appendix Table 12 already evaluates the original directions on seven additional BeaverTails categories and reports an increase in overall DSR from 94.07% to 95.57%. We also evaluate seven external jailbreak attack datasets in Appendix Table 13, where CAM-Steer achieves 82.00% overall DSR, outperforming No Steering by 5.86 percentage points and the strongest baseline by 4.71 percentage points. In this experiment, the steering directions are newly constructed for the seven attack families.

In addition, we conducted a zero-shot evaluation on the same seven jailbreak attack families. For each backbone, we kept the original BeaverTails directions, prototypes, angle budgets, intervention layer, and risk calibration fixed. Under this frozen configuration, overall DSR on the seven jailbreak datasets increased from 76.14% to 77.57% on Qwen, from 30.57% to 33.00% on Llama, and from 38.71% to 39.43% on Gemma. We will also include these new results in the revised manuscript.

AuthorsOpenReview ↗

Updated Manuscript and Follow-up

Dear Reviewer,

Thank you again for your constructive feedback. We have uploaded a revised manuscript incorporating the additional experiments and clarifications discussed in our rebuttal. The corresponding revisions are highlighted in red for ease of reference.

The main updates addressing your concerns can be found in:

  1. Section 5.2 (Table 4) and Appendix D.2 (Tables 14–15): Multi-label risk-score evaluation and harmful leakage analysis.
  2. Section 4.2 and Appendix F (Table 23): Zero-shot transfer evaluation using frozen BeaverTails configurations across three backbones.

We hope these revisions address your concerns and would greatly appreciate any further feedback or questions you may have.

Best regards, The Authors

Thank you for the clarification and the additional analyses. After considering the response, I am updating my rating from 2 (weak reject) to 3 (weak accept).

First, thank you for pointing out the comparisons with MAT-Steer, K-Steering, and ORBIT-R/B in the submitted paper. These comparisons address my question about evaluating existing multi-attribute and rotation-based approaches in the same mixed-harm setting. I had not sufficiently credited this baseline coverage in my initial assessment. I also appreciate the clarified positioning: a category-adaptive steering mechanism, rather than a new steering paradigm.

Second, the multi-label detection and category-specific leakage analyses strengthen the evidence that the risk estimates provide useful information when harms co-occur. The detection results also make the limitations clearer: the estimator does not consistently recover every active category. I therefore view these results as supporting useful category-aware control, rather than complete category identification.

I also find the multi-category comparison with Top-1 Selection and Uniform Weights informative. The additional improvement from the full method on Llama provides more direct evidence that risk-weighted composition contributes beyond selecting a single direction.

My main remaining reservation concerns generalization. The frozen-configuration experiments are valuable, but the gains on external attacks are modest, and several attack families remain poorly defended. These results should be distinguished from experiments that reconstruct directions for the new attack families, and should not be presented as establishing broad robustness.

Overall, I already found the experimental core credible in my initial review. The clarified baseline coverage, stronger component analysis, and more precise positioning now make the case for acceptance more convincing. I support weak acceptance, while encouraging the authors to keep the generalization claims bounded by the reported results.

Thank you for reviewing our responses and for raising your score. Your suggestions have really helped us better position CAM-Steer and strengthen the experimental evidence for its key components. We sincerely appreciate your time and constructive feedback, and we will ensure that our contributions and limitations are clearly presented in the final version.

Reviewer_DBT5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CAM-Steer to address multiple co-occurring harm categories within a single prompt. It uses safe and unsafe prototypes to estimate category-wise risks, forms a weighted combination of safety directions in the tangent space, and applies risk-gated spherical steering with an adaptive rotation angle. The method requires no model parameter updates and applies a single intervention during prefill. The paper evaluates its safety and utility on three models.

Critical strengths
  1. The problem has practical relevance. A single prompt may involve multiple harm categories. The paper addresses this by detecting category-wise risks from the input and using these estimates to compose safety directions and adjust intervention strength.

  2. The method is relatively lightweight, and the experiments cover a broad range of settings. The method requires no model parameter updates and intervenes only once during prefill. The experiments cover three models, multiple harm categories, and utility benchmarks. Overall, the results show improved safety with low inference overhead.

Critical weaknesses
  1. The additional challenges posed by co-occurring harm categories need clearer experimental evidence. The paper takes harm co-occurrence as its main motivation. However, as (K) increases, No Steering and several other methods do not consistently perform worse, and some results even improve. Since different values of (K) correspond to different samples, it is still difficult to tell whether the challenges in these experiments come from harm co-occurrence or from differences between the samples themselves. This also leaves it unclear which specific challenges arising from co-occurrence the proposed design changes address.

  2. Some ablations are missing. Some design choices do seem reasonable, but ablation studies are still necessary. For example, normalizing each category direction separately after projecting it onto the tangent space needs to be evaluated separately. In addition, detector recall should be evaluated to check whether all relevant harm categories can be detected in multi-category prompts.

  3. Counterintuitive experimental results remain unexplained. Compared with Single Detector and Direction, Top-1 Selection already provides considerable gains and comes close to the full method on Qwen and Gemma. These results suggest that using multiple detectors may be a particularly important part of the contribution. However, the paper does not fully explain this finding or clarify what additional gains the weighted composition of multiple directions provides on top of this.

Questions for authors
  1. Could the authors further explain what additional challenges co-occurring harm categories pose for existing activation steering methods and how these challenges are reflected in the experiments? It would also be helpful to explain why No Steering and several other methods do not consistently perform worse as (K) increases, and sometimes even improve, taking into account the sample composition of the different (K) groups. This would help clarify the motivation and the need for the proposed design choices.

  2. Could the authors add an ablation of the tangent-space processing, comparing “normalizing each category direction separately before composition” with “normalizing only after composition,” while keeping the tangent-space projection, risk weights, and rotation angles unchanged? It would also be helpful to report detector recall at the threshold used to trigger steering, to assess whether any harm categories are missed in multi-category prompts.

  3. The results of Single Detector and Direction and Top-1 Selection suggest that using multiple detectors may be a particularly important part of the contribution. Could the authors further explain this finding and clarify the respective roles of multiple detectors and the weighted composition of multiple directions?

5 comments
AuthorsOpenReview ↗

Response to Reviewer DBT5 (1/2)

W1 & Q1: Challenges of co-occurring harms and DSR trends across K

A1: We appreciate the reviewer’s suggestion to clarify the challenges posed by co-occurring harm categories. Unlike single-category settings, multi-category defense requires addressing multiple harm categories within the same prompt. As illustrated in Figure 1 and Appendix G, steering toward safety in one category may leave harmful content from another unaddressed. However, binary DSR alone is insufficient to fully evaluate multi-category defense, as it treats responses leaking one or multiple harm categories equally as unsuccessful defenses. In addition, certain sensitive harm categories may sometimes trigger refusal of the entire prompt, even when other categories alone would not, thereby increasing DSR even as K increases. For instance, Table R1 shows that under No Steering, prompts involving both Financial and Violence achieve 74.00% DSR, closer to the higher DSR of Financial alone (77.59%) , compared with 65.00% for Violence alone. The response-level leakage analysis in Table R2 provides evidence of incomplete category coverage under existing steering methods. CAM-Steer addresses this challenge by adaptively composing safety directions based on category-wise risk scores, and Table R2 also shows its effectiveness in reducing harmful leakage.

Table R1. DSR (%) under No Steering for single-category and co-occurring harm prompts.

Harm categoriesFirst category only (K = 1)Violence only (K = 1)Both categories (K = 2)
Drug and Violence67.8665.0068.00
Financial and Violence77.5965.0074.00

Table R2. Response-level harmful category leakage across different backbones and steering methods.

BackboneMethodExactly oneExactly twoThree or more
Qwen3-8BCAST122522
Qwen3-8BCAM-Steer7176
Llama-3.1-8B-InstructCAA112729
Llama-3.1-8B-InstructCAM-Steer122120
Gemma-2-9B-ITAdaSteer158
Gemma-2-9B-ITCAM-Steer543
AuthorsOpenReview ↗

Response to Reviewer DBT5 (2/2)

W2 & Q2: Directional normalization and category detection

A2: Thank you for the suggestion to isolate the normalization operation. CAM-Steer normalizes each direction after tangent-space projection so that its projected norm does not alter the contribution assigned by its risk weight. In the ablation, we normalized only after composition, keeping the tangent-space projection, risk weights, and rotation angles fixed. On Qwen3-8B, this change reduced average DSR by 2.00 percentage points, from 93.79% to 91.79% (Table R3).

We then evaluate the risk estimator on prompts containing multiple active harm categories. Table R4 reports micro recall of 78.96%, 73.75%, and 76.30% on Qwen, Llama, and Gemma. We also note that incomplete category detection does not necessarily lead to unsuccessful defense, as related safety directions may share representational components that help suppress harmful behaviors associated with missed categories. We will include both the ablation and the detection results in the revised manuscript.

Table R3. Individual normalization and global direction ablations on Qwen3-8B.

VariantAverage DSR (%)Change from Full (pp)
Full CAM-Steer93.790.00
Without individual normalization91.79-2.00
Category-wise Risk Estimation + Global Direction92.14-1.65

Table R4. Detector recall (%) on the multi-category prompts under the steering category activation criterion. Only the seven modeled harm categories are evaluated.

BackboneMicro recallAll-category recall
Qwen3-8B78.9673.25
Llama-3.1-8B-Instruct73.7566.08
Gemma-2-9B-IT76.3068.95

W3 & Q3: Contributions of risk estimation and composition

A3: We agree that the strong performance of Top-1 Selection in Table 3 highlights the importance of category-wise risk estimation. This is expected, as Table 3 evaluates single-category prompts, where the highest-risk direction alone can already provide much of the required protection.

To clarify the respective contributions, we conducted two ablations. As shown in Table R3, category-wise risk estimation improves defense even when using a single global direction. More importantly, to assess the benefit of weighted composition, we compared the full method with Top-1 Selection on multi-category prompts (Table R5). These results support complementary roles for the two components: category-wise risk estimation assesses the risk associated with each harm category, while adaptive direction composition combines the corresponding safety directions based on these risks. We will clarify this distinction in the revised analysis.

Table R5. Direction composition ablation on multi-category prompts using Llama-3.1-8B-Instruct. DSR (%).

MethodK=2K=3K=4K=5Overall
Full CAM-Steer85.0085.5092.0092.8687.74
Top-1 Selection82.5080.5083.5085.7182.32
Uniform Weights83.5084.5089.0089.2985.83
AuthorsOpenReview ↗

Updated Manuscript and Follow-up

Dear Reviewer,

Thank you again for your constructive feedback. We have uploaded a revised manuscript incorporating the additional experiments and clarifications discussed in our rebuttal. The corresponding revisions are highlighted in red for ease of reference.

The main updates addressing your concerns can be found in:

  1. Section 4.2 and Appendix D.2 (Tables 13–15): Additional analysis of the challenges arising from co-occurring harm categories, DSR trends across K, and harmful category leakage.
  2. Section 5.2 (Table 4): Multi-label harm category detection, including detector recall.
  3. Appendix B.2 (Table 7) and Appendix D.2 (Table 16): Additional ablations examining individual direction normalization and the respective contributions of category-wise risk estimation and adaptive direction composition.

We hope these revisions address your concerns and would greatly appreciate any further feedback or questions you may have.

Best regards, The Authors

Reviewer_TXz6OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies inference-time activation steering for harmful prompts involving multiple safety categories. It proposes CAM-Steer, which constructs safe and unsafe prototypes and a safety direction for each harm category, estimates category-wise risk scores from the final prompt-token representation at a selected transformer layer, and uses these scores to combine projected safety directions and determine the strength of a norm-preserving spherical rotation. The authors evaluate the method on three 8–9B language models and seven BeaverTails harm categories, including naturally multi-labeled prompts, and report higher average defense success rates than the evaluated activation-steering baselines. Additional experiments examine utility preservation, component ablations, risk-score selectivity, direction-construction stability, unseen categories, jailbreak attack families, and inference latency.

Critical strengths
  1. The paper addresses an important and underexplored safety problem. Harmful prompts can involve multiple categories simultaneously, while many activation-steering methods operate with a single direction or intervention objective. Framing safety steering as an input-dependent multi-category problem is relevant to practical LLM deployment.

  2. CAM-Steer is technically coherent and clearly motivated. The method combines category-wise risk estimation, adaptive direction composition, and a spherical update that preserves the hidden-state norm. The mathematical formulation is understandable, and the relationship between the risk scores, direction weights, and intervention strength is explicit.

  3. The main BeaverTails evaluation is broad within its chosen setting. The paper evaluates three model families, seven harm categories, naturally multi-labeled prompts, and nine activation-steering baselines. CAM-Steer achieves the highest reported average DSR across the three backbones in both individual- and multi-category evaluations.

  4. The ablation study includes useful controlled comparisons. In addition to removing individual components, the paper compares spherical steering with renormalized and displacement-matched additive updates. These controls provide stronger evidence for the spherical update than a comparison with conventional additive steering alone.

  5. The paper includes meaningful robustness, utility, and efficiency analyses. Construction, calibration, and evaluation data are reportedly disjoint; the authors examine direction-construction seeds, sample size, intervention-layer sensitivity, unseen categories, utility benchmarks, and latency. These analyses improve confidence that the main BeaverTails results are not caused by one construction seed or a large direction-estimation dataset.

Critical weaknesses
  1. Single-evaluator bias. All DSR results rely on one GPT-4.1-mini evaluator without human validation, an independent judge, evaluator-prompt sensitivity analysis, or paired confidence intervals. Because several margins over the strongest baseline are below two percentage points, evaluator-specific preferences could change the ranking. Validation with blinded human annotations and multiple independent judges likely requires substantial new work.

  2. The multi-category evaluation does not directly test category identification. Tables 2 and 10 measure only whether the final response is safe, so a generic refusal receives full credit even if CAM-Steer fails to identify all active categories. Multi-label precision, recall, exact-match accuracy, and per-category response leakage are needed to establish genuine category coordination. This likely requires substantial new work.

  3. Cross-dataset generalization is limited. Direction construction, calibration, and the complete three-backbone comparison all use BeaverTails; although the records are disjoint, they share the same distribution and taxonomy. The unseen-category experiment also remains within BeaverTails, while the jailbreak experiment reconstructs new attack-family directions rather than testing zero-shot transfer. Evaluating the original directions on an external safety dataset likely requires substantial new work.

  4. Model coverage outside BeaverTails is selective. The all-baseline XSTest comparison uses only Llama, while the unseen-category and jailbreak experiments use only Qwen, and no backbone-selection criterion is provided. It is therefore unclear whether these results generalize across model families. Additional backbone experiments are likely required.

  5. A complete safety–utility comparison is missing. The six-benchmark utility evaluation compares CAM-Steer only with No Steering, while the all-baseline utility comparison is limited to XSTest on Llama. The paper therefore does not establish that CAM-Steer provides a better safety–utility tradeoff than prior methods at matched refusal rates. A complete baseline utility evaluation likely requires additional work.

  6. Baseline construction and tuning fairness cannot be verified. The paper states that tunable methods share a calibration split and selection criterion, but does not report each baseline’s construction data, sample counts, search space, selected layer, final steering strength, or tuning budget. Some reported differences could therefore result from unequal data or tuning effort. These details may be resolvable during discussion if the configurations already exist.

  7. Intervention-layer selection is unexplained. The paper uses layers 10, 18, and 20 for Qwen, Llama, and Gemma without describing the candidate layers, selection criterion, or calibration results. The Qwen layer sweep demonstrates robustness but does not explain why layer 10 was selected, and equivalent sweeps are absent for Llama and Gemma. This may be resolvable during discussion if the layer-selection experiments already exist.

  8. Aggregate DSR obscures response-level regressions. DSR reports only the final proportion of safe responses and does not show how often CAM-Steer converts unsafe responses to safe ones or originally safe responses to unsafe ones. It also treats borderline and unsafe responses identically, despite potentially different harm severity. Paired transition counts may be obtainable from existing outputs, while graded severity analysis would require additional annotation.

  9. The final-token representation is insufficiently analyzed. CAM-Steer represents the complete prompt using one final prompt-token vector without comparing mean pooling, attention pooling, other token positions, or multi-token representations. The representation is empirically useful, but its sufficiency and optimality for co-occurring harms are not established. A representation-level ablation would require additional experiments.

Questions for authors
  1. How robust are the reported DSR rankings to evaluator choice?
    Could the authors provide blinded human evaluation and results from at least one independent safety classifier or LLM judge, together with inter-rater agreement and paired confidence intervals? This is particularly important where CAM-Steer’s margin over the strongest baseline is below two percentage points.

  2. Does CAM-Steer correctly identify and address all active categories in multi-category prompts?
    Please report multi-label precision, recall, exact-match accuracy or all-category recall, and per-category harmful leakage. Because a generic refusal receives full DSR credit, Tables 2 and 10 do not distinguish genuine multi-category coordination from general refusal behavior.

  3. How were intervention layers 10, 18, and 20 selected?
    Please report the candidate layers, calibration-only selection criterion, per-layer calibration results, and final selection procedure for all three backbones. How were intervention layers selected for the baselines?

  4. Can the authors provide a matched safety–utility comparison for all baselines?
    Figures 3 and 8 compare only CAM-Steer with No Steering across the six utility benchmarks, while Table 11 includes all baselines only for XSTest on Llama. Please report complete baseline utility results or safety–utility curves at matched refusal rates.

  5. How well do the original BeaverTails-derived directions transfer without reconstruction or recalibration?
    The unseen-category experiment remains within BeaverTails, while the jailbreak experiment constructs new attack-family directions and evaluates only Qwen. Please evaluate the original directions on an external safety dataset across multiple backbones and justify the backbone choices in the appendix experiments.

8 comments
AuthorsOpenReview ↗

Response to Reviewer TXz6 (1/5)

W1 & Q1: Evaluation consistency across judges

A1: We thank the reviewer for raising the question of how reliably DSR captures differences between methods. To test whether evaluator choice affected the ranking, we used Claude Sonnet 4.5 to independently reassess the same responses from CAM-Steer and the strongest baseline in each setting. Under this independent evaluation, CAM-Steer retained its lead in all six settings, with DSR gains ranging from 0.29 to 3.43 percentage points (Table R1). The original GPT-4.1-mini judgments agreed with each of two human annotators in 97.0% of cases, and the annotators agreed with each other in 98.0% (Table R2).

To quantify uncertainty in the original evaluation, we computed paired 95% confidence intervals for the DSR difference between CAM-Steer and the corresponding baseline (Table R1). For both Qwen and Llama, the intervals lay entirely above zero in both settings. Gemma showed smaller gains, consistent with its already high baseline DSR and the limited room for improvement.

Table R1. Independent evaluation using Claude Sonnet 4.5 and paired confidence intervals for the original GPT-4.1-mini evaluation. Δ is the DSR difference between CAM-Steer and the selected strongest non-oracle baseline in percentage points. Confidence intervals are computed from the original paired evaluations.

SettingBackboneBaselineClaude CAM DSR (%)Claude baseline DSR (%)Claude Δ (pp)Original Δ (pp)Original paired 95% CI (pp)
SingleQwen3-8BAdaSteer95.4392.71+2.71+3.07[1.36, 4.79]
SingleLlama-3.1-8B-InstructCAA94.4391.00+3.43+5.57[3.86, 7.29]
SingleGemma-2-9B-ITAdaSteer97.9397.64+0.29+0.21[-0.71, 1.21]
MultiQwen3-8BCAST94.4391.88+2.55+3.50[0.48, 6.53]
MultiLlama-3.1-8B-InstructCAA90.4589.17+1.27+1.75[0.16, 3.50]
MultiGemma-2-9B-ITAdaSteer98.7398.41+0.32+0.32[-0.96, 1.59]

Table R2. Agreement between the original GPT-4.1-mini evaluator and two human annotators in 100 random prompts.

Evaluator pairAgreement (%)Cohen’s κ
GPT-4.1-mini and Human 1970.753
GPT-4.1-mini and Human 2970.713
Human 1 and Human 2980.790
AuthorsOpenReview ↗

Response to Reviewer TXz6 (2/5)

W2 & Q2: Category-wise assessment on mixed-harm prompts

A2: We agree that overall DSR in Tables 2 and 10 does not directly measure whether the risk estimator identifies every active harm category. Figure 5 and Appendix Figure 9 already examine category selectivity and source-category coverage. To assess category identification more directly, we also evaluated multi-label harm detection on the prompts containing two to five active harm categories. Table R3 reports micro F1 scores of 77.22%, 70.58%, and 67.82% for Qwen, Llama, and Gemma, respectively. More importantly, a response can still be safe even when the estimator misses an active category. CAM-Steer uses continuous risk scores to weight the steering directions. Safety directions for related harm categories may share representational components, so steering along the detected directions could suppress overlapping harmful behaviors even when some active categories go undetected.

To assess the final responses, we also evaluated category-specific harmful leakage (Tables R4 and R5). Table R4 reports the number of harm categories leaked per response, and Table R5 reports the leakage rate for each category. We will include the detailed results and evaluation procedure in the paper.

Table R3. Multi-label harm category detection on multi-category prompts for each backbone. Metrics are percentages over the seven modeled harm categories. Exact match requires an identical predicted and ground-truth category set. All-category recall requires all active modeled categories to be detected.

BackboneMicro precisionMicro recallMicro F1Exact matchAll-category recall
Qwen3-8B75.5678.9677.2247.6173.25
Llama-3.1-8B-Instruct67.6773.7570.5834.7166.08
Gemma-2-9B-IT61.0376.3067.8229.7868.95

Table R4. Response-level harmful category leakage across different backbones and steering methods.

BackboneMethodExactly oneExactly twoThree or more
Qwen3-8BCAST122522
Qwen3-8BCAM-Steer7176
Llama-3.1-8B-InstructCAA112729
Llama-3.1-8B-InstructCAM-Steer122120
Gemma-2-9B-ITAdaSteer158
Gemma-2-9B-ITCAM-Steer543

Table R5. Harmful leakage rate (%) for each of the seven modeled categories under the category-specific evaluation.

Harm categoryQwen CASTQwen CAMLlama CAALlama CAMGemma AdaSteerGemma CAM
Hate5.771.927.375.451.601.92
Drug5.632.827.042.820.000.00
Financial6.413.216.417.691.921.92
Privacy8.066.454.844.843.231.61
Self-harm0.000.000.000.000.000.00
Sexual4.880.004.882.440.000.00
Violence3.271.794.172.380.890.30
AuthorsOpenReview ↗

Response to Reviewer TXz6 (3/5)

W3 & Q5: Transfer with frozen BeaverTails configurations

A3: We used Qwen for the attack-family experiment in Appendix Table 13 because it was our primary backbone. To assess zero-shot transfer across backbones, we now evaluate Qwen, Llama, and Gemma on seven external jailbreak attack datasets, with 100 prompts per family. For each backbone, we retain the original BeaverTails directions, prototypes, angle budgets, intervention layer, and risk calibration. Across the seven families, CAM-Steer improved overall DSR over No Steering by 1.43 percentage points on Qwen, 2.43 on Llama, and 0.71 on Gemma (Table R6). Surprisingly, for some attack families, DSR remained low under both methods (e.g., Llama achieved 0% DSR both with and without steering on AIM). This suggests that the effectiveness of zero-shot transfer is also influenced by the backbone's existing safety capabilities, as the original steering directions provide limited improvements on attacks where the model itself struggles. We will clarify the distinction between these evaluations and include the new transfer results in the revised manuscript.

Table R6. DSR (%) on seven external safety datasets using frozen BeaverTails directions.

BackboneMethodAIMAutoDANCipherGCGJailbrokenMultilingualReNeLLMOverall
Qwen3-8BNo Steering8290956885644976.14
Qwen3-8BCAM-Steer8289977488635077.57
Llama-3.1-8B-InstructNo Steering037677530530.57
Llama-3.1-8B-InstructCAM-Steer077580621633.00
Gemma-2-9B-ITNo Steering02482876231338.71
Gemma-2-9B-ITCAM-Steer03079896231339.43

W4: Backbone coverage in auxiliary experiments

A4: We appreciate the reviewer’s concern about backbone coverage in the auxiliary evaluations. We primarily use Qwen3-8B for supplementary analyses, while the main safety experiments cover all three backbones. The all-baseline XSTest comparison was conducted on Llama to investigate why CAA, despite its relatively simple steering mechanism, achieves higher DSR than several more sophisticated baselines. In particular, as explained in Appendix D.4, the XSTest results suggest that CAA's relatively strong safety performance comes partly at the cost of over-refusal.

W5 & Q4: Safety–utility comparison across methods

A5: We appreciate the reviewer’s suggestion to extend the utility comparisons to all baselines. We now evaluate all methods on six utility benchmarks for Llama under their selected configurations in table R7. This is consistent with our objective of preserving general utility while improving safety, as CAM-Steer achieves the highest overall DSR in the main evaluation without sacrificing performance on these utility benchmarks. We will include the complete comparison in the revised manuscript.

Table R7. Utility results (%) on Llama-3.1-8B-Instruct for No Steering, nine baselines, and CAM-Steer. Higher values indicate better performance.

MethodXSTestAlpacaEvalMATHGSM8KMMLUHumanEval
No Steering958940886586
CAA854640904786
SafeSteer958540886588
CAST986524342144
AdaSteer996026504632
AlphaSteer995930666052
MAT-Steer965722626468
K-Steering874736845884
ORBIT-R956342906486
ORBIT-B966440826784
CAM-Steer968940886686

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair and Reviewers,

We sincerely thank the reviewers for their thoughtful and detailed feedback, as well as for recognizing the key strengths of our work. We are happy to read that the reviewers found our work to address an important and underexplored safety problem, with a coherent and lightweight design and strong empirical evidence supporting the effectiveness of CAM-Steer in the evaluated setting.

To facilitate a fair assessment, we present a summary of the current rebuttal status below:

ReviewerConcern / RequestOur Response
TXz6Evaluation consistency across judges (W1 & Q1)Validated results using Claude Sonnet 4.5, human annotations, and paired 95% confidence intervals.
Category-wise assessment on mixed-harm prompts (W2 & Q2)Added multi-label detection and category-specific harmful leakage analyses.
Transfer with frozen BeaverTails configurations (W3 & Q5)Evaluated seven external jailbreak attack datasets using frozen BeaverTails configurations.
Backbone coverage in auxiliary experiments (W4)Clarified backbone choices for auxiliary analyses.
Safety–utility comparison across methods (W5 & Q4)Extended utility comparisons to all baselines across three backbones.
Baseline construction and calibration protocol (W6)Clarified the shared construction pool, baseline configurations, tuning budgets, and selected settings.
Intervention-layer selection protocol (W7 & Q3)Added layer selection results across candidate layers 10–22.
Paired response outcomes and evaluation criteria (W8)Added paired safety transition analysis.
Prompt representation alternatives (W9)Added prompt-representation ablations.
DBT5Challenges of co-occurring harms and DSR trends (W1 & Q1)Clarified DSR trends across different harm-category combinations and added category-specific harmful leakage analyses.
Directional normalization and category detection (W2 & Q2)Added direction-normalization ablations and detector recall results.
Contributions of risk estimation and composition (W3 & Q3)Added global-direction and multi-category direction-composition ablations.
CuzpPositioning against existing steering approaches (W1 & Q1)Clarified comparisons with MAT-Steer, K-Steering, and ORBIT-R/B, and refined the positioning of CAM-Steer.
Multi-label risk-score behavior (W2 & Q2)Added multi-label detection and category-specific harmful leakage analyses.
Evaluation scope across taxonomies (W2 & Q3)Evaluated seven external jailbreak attack datasets using frozen BeaverTails configurations.

In particular, all three reviewers (TXz6, DBT5, and Cuzp) raised their ratings from 2 to 3 after considering our responses and additional experiments. We greatly appreciate their time and constructive feedback, which has truly helped us strengthen the experimental evidence and clarify the contributions of CAM-Steer.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.