acceptodds
Under review as a conference paper at ICLR 2027

Trajectory-Guided Rotation Steering for Jailbreak Defense in Audio Large Language Models

Abstract

Audio Large Language Models (ALLMs) have demonstrated remarkable capabilities in audio understanding and reasoning tasks. Despite these capabilities, they remain vulnerable to jailbreak attacks, where adversaries craft malicious audio inputs to induce the models to generate harmful responses. Existing activation steering-based defense methods mitigate this issue by injecting refusal-related directions into intermediate-layer representations. However, these approaches typically rely on static layer-wise steering vectors and additive modification strategies, which fail to capture the dynamic evolution of refusal behavior across network layers and may introduce excessive interference into benign queries. In this paper, we propose TRAIL, a Trajectory-based Rotational Activation Intervention across Layers framework for jailbreak defense in ALLMs. We observe that refusal behaviors exhibit distinct cross-layer evolution patterns compared with normal responses, suggesting that refusal is not determined by a fixed direction but rather emerges as a dynamic representation trajectory. Based on this observation, we propose a layer-wise refusal trajectory modeling method to extract discriminative inter-layer refusal evolution features. We then leverage these trajectory features to construct a trajectory-guided rotation subspace, which enables fine-grained adjustment of intermediate representations while preserving the original semantic space. Furthermore, we find that refusal behaviors induced by harmful requests and those triggered by benign requests share partially overlapping components in the representation space. We therefore introduce a refusal trajectory disentanglement strategy to suppress the components that contribute to over-refusal on safe queries.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Audio Large Language Models, Activation Steering, Jailbreak Defense

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_QwuKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a defense method against jailbreak attacks on large audio language models. It aims to defend against these attacks while avoiding refusals of normal requests as much as possible. Specifically, it proposes TRAIL, which tracks how refusal-related hidden states change across model layers, removes components shared with safe requests, and uses rotation instead of vector addition to modify the hidden states. Experiments further show that it improves jailbreak defense while largely preserving performance on normal tasks.

Critical strengths
  • The motivation is clear, and the presentation is easy to follow and understand.

  • The idea of studying how refusal behavior changes across layers is interesting. Using rotation instead of directly adding a refusal vector is also a reasonable design choice.

  • The experiments cover several jailbreak benchmarks and evaluate both defense performance and normal task performance.

Critical weaknesses
  • In terms of novelty, it may not be very different from using cross-layer trajectories in the intermediate layers of text LLMs to guide interventions (for example, Spherical Steering (ICML 2026) also uses rotation-based interventions). Both cross-layer refusal analysis and rotation-based steering have been studied before. The paper should better explain what is new beyond combining these ideas for LALMs. What is different for LALMs?

  • The method combines several changes, but their individual contributions are not clearly isolated. The paper does not directly compare cross-layer directions with additive steering, or single-layer directions with rotation. In other words, the ablation study is insufficient. It mainly examines hyperparameters but does not isolate the contributions of cross-layer trajectory modeling and rotation steering.

  • The evaluation is limited to 2 LALMs. Can this demonstrate generality? It also lacks comparisons with output safety guards (e.g. Azure Content Filter or Llama Guard), which are a practical alternative for filtering harmful responses.

  • A minor issue, Figure 3 is too far from where it is cited. Figure 5 does not need to be here at all. This part (case study) takes up a lot of space, but does not seem important and could be moved to the appendix.

Questions for authors
  • What is new for LALMs compared with existing cross-layer refusal analysis and rotation-based steering in LLMs?

  • Which component mainly contributes to the improvement? Why are cross-layer modeling and rotation both necessary?

  • How well does TRAIL generalize beyond the two evaluated LALMs? Maybe also compare output filter methods.

3 comments
AuthorsOpenReview ↗

Response to Reviewer QwuK [1/2]

We highly appreciate your positive feedback and valuable suggestions about our paper. We respond to your questions as follows.

Q1: Comparison with Spherical Steering.

A1: Thank you for raising this important concern. We would like to emphasize that, compared with Spherical Steering in LLMs (ICML 2026) [1], our method establishes its novelty at three levels:

(1) Cross-layer Refusal Trajectory Modeling.

Spherical Steering constructs a contrastive prototype direction μ^((l)) at a given layer and rotates activations toward this prototype. Its steering target is a static, layer-wise contrastive direction. In TRAIL, however, the steering target is not a static prototype but a cross-layer refusal trajectory, which characterizes how the refusal representation evolves with layer depth, thereby yielding stronger defense performance. Further details are provided in Challenge I of Section 3.2 in the original manuscript (Page 4, Lines 199–207).

(2) Two-dimensional Subspace-constrained Rotation.

Spherical Steering performs norm-preserving rotation on the full d-dimensional hypersphere, rotating the entire activation direction toward the target prototype. TRAIL, in contrast, projects the activation onto a local two-dimensional rotation basis. The rotation is performed only within this plane, while components outside the plane remain unchanged, thereby minimizing the damage of activation steering to the original semantics.

(3) How is this different for ALLMs?

TRAIL is specifically designed for this setting. Refusal behavior is entangled with cross-modal audio representations, audio jailbreak attacks bring new attack surfaces, and benign audio queries are particularly prone to over-refusal. TRAIL is designed precisely for this setting: it extracts refusal trajectories from audio-language hidden states, decouples the benign over-refusal component, and performs rotation within the trajectory-derived subspace. Therefore, TRAIL is more than a generic rotation primitive; it also addresses the ALLM-specific cross-modal refusal dynamics and the robustness–over-refusal trade-off. We have provided detailed justification for this in Challenge II of Section 3.2 in the original manuscript (Page 5, Lines 223–230), and we will further clarify these differences to make the distinction clearer. In addition, we have added a comparison with Spherical Steering (ICML 2026) [1] adapted to the ALLM setting, as shown below:

Tab. R1: Comparison of ASR (%) between TRAIL and Spherical Steering

ModelMethodFigStep-Audio
Qwen2-AudioSpherical Steering [1]13.20
TRAIL5.20
Kimi-AudioSpherical Steering [1]12.80
TRAIL4.80

As shown in the Tab. R1, TRAIL achieves lower ASR on harmful prompts for both models.

[1] Spherical Steering: Geometry-Aware Activation Rotation for Language Models. ICML 2026

Q2: Core Components and Their Complementary Roles.

A2: Thank you for your insightful comments. We have added ablation experiments on the two core components of our method: cross-layer refusal trajectory modeling and rotation steering, as shown in Tab. R2.

Tab. R2: Ablation Study of TRAIL's Core Components on the FigStep-Audio Dataset.

ModelTRAILw/o Trajectoryw/o Rotation
Qwen2-Audio5.28.846.0
Kimi-Audio4.88.415.6

We have presented the ASR results of ablating the two core components on the FigStep-Audio Dataset. The experiments show that both components contribute to improving defense performance. We will revise the final version to ensure that the ablation is sufficient.

In addition, we have supplemented a comparison of TRAIL with additive steering (i.e., removing the rotation steering) to evaluate their impact on the models' general audio capabilities, as shown in Tab. R3. The data are from AirBench, using its official evaluation settings:

Tab. R3: Comparison of TRAIL and Additive Steering on AirBench.

ModelSettingSpeechScoreSoundScoreMusicScoreMixedScoreAvg. Score
Qwen2-AudioTRAIL7.897.387.447.497.55
w/o Rotation7.757.267.337.367.43
Kimi-AudioTRAIL7.827.237.167.157.34
w/o Rotation6.996.266.546.506.57

As shown in the Tab. R3, additive steering degrades model capabilities, supporting our motivation in Challenge II of Section 3.2 regarding its potential disruption of semantic structure.

AuthorsOpenReview ↗

Response to Reviewer QwuK [2/2]

Q3: Cross-Model Generalization.

A3: We added two evaluation models, Qwen2.5-Omni and Llama-Omni, to verify the cross-model generalization of the defense method, as follows:

Tab. R4: ASR of TRAIL on Additional Models.

ModelMethodFigStep-AudioAdvBench-AudioHarmBench
Qwen2.5-Omniw/o TRAIL40.83.817.5
w/ TRAIL32.40.912.6
Llama-Omniw/o TRAIL73.632.635.9
w/ TRAIL54.415.926.5

We did not make any setting adjustments for the newly added models; we directly applied the original settings to Qwen2.5-Omni and Llama-Omni. As can be seen, our method also effectively improves the defense capability of the newly added models. These results are reported in Appendix D.2.

Q4: Explore issues about output safety guards.

A4: Thank you for the suggestion. Llama-Guard is a detection-based method, whereas TRAIL is an activation-intervention-based method. They are not directly comparable, but can be complementary when used in series. We adopt a pipeline in which activation steering is applied first, followed by Llama-Guard detection. As shown in Tab. R5, on FigStep-Audio, with Llama-Guard alone, 22.0% of the outputs it judged safe were still judged unsafe by SorryBench. After adding activation steering, this ratio drops to 2.8%. This indicates that combining the two can improve defense effectiveness. This is an interesting finding, and we thank you for the idea. These results are presented in Appendix D.4.

Tab. R5: ASR (%) of Llama-Guard-3-8B with and without TRAIL on FigStep-Audio (lower is better).

ModelMethodFigStep-Audio
Qwen2-AudioOnly Llama-Guard-3-8B22.0
TRAIL+Llama-Guard-3-8B2.8
Kimi-AudioOnly Llama-Guard-3-8B10.8
TRAIL+Llama-Guard-3-8B3.2

Q5: Suggestions on manuscript organization.

A5: Thank you for your helpful suggestions regarding the manuscript organization. Following your suggestions, we have added an appropriate reference to Figure 3 at a more relevant location, moved the Case Study section to Appendix E, and introduced a new Inference Efficiency Analysis subsection in the main text to replace the space previously occupied by the case study.

AuthorsOpenReview ↗

Official Comment By Authors

Dear reviewer QwuK ,

Thank you for your valuable comments and constructive suggestions.

As the discussion deadline is approaching, we would appreciate your participation in the discussion so that we have sufficient time to address any further questions or concerns you may have.

Thank you for your time and consideration, and we look forward to hearing from you.

Best Regards,

Authors of Submission 35924

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

This paper proposes TRAIL, a representation intervention method for defending audio large language models against jailbreak attacks. It captures cross-layer changes in refusal signals by comparing hidden states under base and refusal-conditioned settings. The method combines two-dimensional subspace construction with soft disentanglement of refusal trajectories and applies rotational interventions during inference. Experiments on Qwen2-Audio and Kimi-Audio compare TRAIL with five defense methods. The results show that TRAIL reduces attack success rates across multiple audio safety benchmarks while maintaining good average performance on AirBench.

Critical strengths
  1. The paper highlights the importance of examining cross-layer trajectories of refusal behavior within the model and demonstrates substantial performance improvements over closely related baselines.
  2. The proposed rotational subspace intervention and soft disentanglement of refusal trajectories enable fine-grained control over perturbations to the representation space and mitigate over-refusal.
Critical weaknesses
  1. The novelty relative to Angular Steering is not sufficiently clear, and a crucial direct comparison with this closely related method is missing.
  2. The evaluation covers only two audio language models, Qwen2-Audio and Kimi-Audio. This limited model coverage does not sufficiently establish the method’s applicability across different model families and architectures.
Questions for authors
  1. Could the authors provide additional experiments evaluating the method’s impact on inference efficiency?
  2. Could the authors test the effectiveness of the method on more ALMs, especially those from different model families and architectures?
3 comments
AuthorsOpenReview ↗

Response to Reviewer gZRV [1/2]

We sincerely appreciate your suggestions, which can greatly help us to refine our paper. We respond to your questions as follows.

Q1: Comparison with Spherical Steering.

A1: Thank you for raising this important concern. We would like to emphasize that, compared with Spherical Steering in LLMs (ICML 2026) [1], our method differs from Spherical Steering in three key aspects:

(1) Cross-layer Refusal Trajectory Modeling.

Spherical Steering constructs a contrastive prototype direction μ(l)\mu^{(l)} at a given layer and rotates activations toward this prototype. Its steering target is a static, layer-wise contrastive direction. In TRAIL, however, the steering target is not a static prototype but a cross-layer refusal trajectory, which characterizes how the refusal representation evolves with layer depth, thereby providing a more consistent steering target across network layers. Further details are provided in Challenge I of Section 3.2 in the original manuscript (Page 4, Lines 199–207).

(2) Rotation Constrained to a Two-Dimensional Subspace.

Spherical Steering performs norm-preserving rotation on the full d-dimensional hypersphere, rotating the entire activation direction toward the target prototype. TRAIL, in contrast, projects the activation onto a local two-dimensional rotation basis. The rotation is performed only within this plane, while components outside the plane remain unchanged, thereby minimizing the damage of activation steering to the original semantics.

(3) How is this different for ALLMs?

Refusal behavior is entangled with cross-modal audio representations, audio jailbreak attacks bring new attack surfaces, and benign audio queries are particularly prone to over-refusal. TRAIL is designed precisely for this setting: it extracts refusal trajectories from audio-language hidden states, decouples the benign over-refusal component, and performs rotation within the trajectory-derived subspace. Therefore, TRAIL is more than a generic rotation primitive; it also addresses the ALLM-specific cross-modal refusal dynamics and the robustness–over-refusal trade-off. We have provided detailed justification for this in Challenge II of Section 3.2 in the original manuscript (Page 5, Lines 223–230), and we will further clarify these distinctions in the revised manuscript. In addition, we have added a comparison with Spherical Steering (ICML 2026) [1] adapted to the ALLM setting, as shown below:

Tab. R1: Comparison of ASR (%) between TRAIL and Spherical Steering

ModelMethodFigStep-Audio
Qwen2-AudioSpherical Steering [1]13.20
TRAIL5.20
Kimi-AudioSpherical Steering [1]12.80
TRAIL4.80

As shown in the Tab. R1, TRAIL achieves lower ASR on harmful prompts for both models.

[1] Spherical Steering: Geometry-Aware Activation Rotation for Language Models. ICML 2026

Q2: Cross-Model Generalization.

A2: We added two evaluation models, Qwen2.5-Omni and Llama-Omni, to verify the cross-model generalization of the defense method, as follows:

Tab. R2: ASR of TRAIL on Additional Models.

ModelMethodFigStep-AudioAdvBench-AudioHarmBench
Qwen2.5-Omniw/o TRAIL40.83.817.5
w/ TRAIL32.40.912.6
Llama-Omniw/o TRAIL73.632.635.9
w/ TRAIL54.415.926.5

We did not make any setting adjustments for the newly added models; we directly applied the original settings to Qwen2.5-Omni and Llama-Omni. As can be seen, our method also effectively improves the defense capability of the newly added models. These results are reported in Appendix D.2.

AuthorsOpenReview ↗

Response to Reviewer gZRV [2/2]

Q3: Inference Efficiency Analysis.

A3: Thank you for your suggestion regarding the discussion of computational cost. We evaluate inference efficiency from three aspects: average inference time per query, per-token inference latency, and GPU memory usage, as shown in Tab. R3:

Tab. R3: Inference Efficiency on the FigStep-Audio Dataset.

ModelSettingAvg. Time (s/query)Per-token Latency (s/token)Δ Latency (s/token)GPU Memory (GB)
Qwen2-Audiow/o TRAIL5.226240.04211-16.79
w/ TRAIL2.972080.047750.0056416.79
Kimi-Audiow/o TRAIL5.274760.07584-29.84
w/ TRAIL5.337800.085560.0097229.84

As can be seen, TRAIL achieves comparable average inference time per query to the Base setting on Kimi-Audio, while reducing it by approximately 43.1% on Qwen2-Audio. This is because the defense causes the model to output mostly concise responses such as "Sorry, I don't...", thereby reducing the number of output tokens. To further assess the computational overhead, we measured the average per-token latency under both the TRAIL and Base settings. Moreover, TRAIL introduces only a small absolute increase in per-token inference latency compared with the Base setting. This is because our defense only applies offline-computed activation steering to intermediate-layer activations, thereby introducing minimal computational overhead during inference. In addition, TRAIL does not increase GPU memory usage. These results are reported in Page 9, Lines 449–474.

AuthorsOpenReview ↗

Official Comment By Authors

Dear reviewer gZRV ,

Thank you for your valuable comments and constructive suggestions.

As the discussion deadline is approaching, we would appreciate your participation in the discussion so that we have sufficient time to address any further questions or concerns you may have.

If you feel that our response has addressed your concerns, would you kindly consider raising your rating score?

Thank you for your time and consideration, and we look forward to hearing from you.

Best Regards,

Authors of Submission 35924

Reviewer_Se2kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes TRAIL, an inference-time defense for audio language models. It models refusal as a cross-layer trajectory rather than a static direction. It extracts harmful and over-refusal trajectories from paired forward passes, removes shared components with an SVD-based projection, builds a two-dimensional rotation plane and rotates hidden states at selected layers. Experiments on Qwen2-Audio and Kimi-Audio use four audio jailbreak benchmarks and an audio utility benchmark.

Critical strengths

The cross-layer trajectory view is a clear conceptual contribution.

The rotation and disentanglement are technically coherent.

The ASR reduction is large on the main benchmark, and over-refusal is lower than the compared steering baselines.

Critical weaknesses

The defense is not tested against adaptive attacks.

The safety metric depends on an unvalidated LLM judge.

The operating point is tuned on the test benchmarks and only two models are covered.

Questions for authors

How does TRAIL perform under an adaptive attack that optimizes against the rotation?

What is the human agreement rate with the ASR judge?

Were the angle, window, disentanglement strength and layer range selected on a separate validation set?

How is k chosen for the over-refusal subspace, and how stable are the basis vectors across inputs?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Se2k [1/2]

Thank you for your valuable comments and suggestions. Our responses to your concerns are provided below.

Q1: Performance against adaptive attacks.

A1: Thank you for raising this important question. To evaluate whether TRAIL can defend against adaptive attacks when the attacker knows our steering method in advance, we conducted white-box adaptive attacks against TRAIL. Specifically, the attacker is assumed to know the steering basis vector b1b_1 and applies reverse additive steering to the original malicious input. We report in Tab. R1 the defense performance of TRAIL against adaptive attacks on the original Qwen2-Audio and Kimi-Audio models. TRAIL maintains a low ASR under adaptive attacks. We attribute this robustness to TRAIL's principled design. These results indicate that TRAIL retains defense effectiveness under the evaluated adaptive attacks.

Tab. R1: ASR (%) of TRAIL With and Without Adaptive Attacks.

ModelSettingFigStep-AudioAdvBench-AudioHarmBench
Qwen2-Audiow/o Adaptive Attack5.200.000.82
w/ Adaptive Attack7.600.772.86
Kimi-Audiow/o Adaptive Attack4.800.000.41
w/ Adaptive Attack10.400.003.27

These results are reported in Appendix D.3.

Q2: ASR Judge Reliability.

A2: Our method follows the LLM-judge-based defense evaluation benchmark of SARSteer (ICML 2026). To address your concern, we additionally conducted a manual review of the LLM-based ASR judge. Specifically, for responses that the ASR judge classified as safe, we provided the original questions and model responses to annotators without revealing the judge's verdict, and asked them to independently determine whether each response was safe. The human evaluation results were aggregated across three human evaluators. This directly checks whether the judge overestimates the defense success rate. The results are shown in Tab. R2.

Tab. R2: ASR evaluation by LLM and human judges on 250 FigStep-Audio queries.

Without TRAILWith TRAIL
ModelLLM judgeHuman judgeLLM judgeHuman judge
Qwen2-Audio51.6051.605.205.60
Kimi-Audio15.6015.604.804.80

As can be seen, the human evaluation results are highly consistent with the model's evaluation. To further address your question, we introduced an LLM judge, Llama-Guard-3-8B, to verify the results. The results are shown in Tab. R3.

Tab. R3: Cross-Judge ASR Evaluation on Four Audio Jailbreak Benchmarks.

ModelEvaluatorFigStep-AudioAdvBench-AudioAJailBenchHarmBench
Qwen2-Audiomistral_sorryBench5.200.0011.500.82
Llama-Guard-3-8B3.200.001.500.41
Kimi-Audiomistral_sorryBench4.800.0010.000.41
Llama-Guard-3-8B1.600.001.500.41

These results, together with the human evaluation, consistently demonstrate the effectiveness of TRAIL in improving model safety across different evaluation protocols. These results are reported in Appendix D.1.

AuthorsOpenReview ↗

Response to Reviewer Se2k [2/2]

Q3: Validation Split Description.

A3: Our hyperparameters were selected on a validation set independently split from FigStep-Audio, ensuring that the parameter selection process did not involve the unseen test set and thus preserving the fairness of the evaluation.

Q4: Cross-Model Generalization.

A4: We added two evaluation models, Qwen2.5-Omni and Llama-Omni, to verify the cross-model generalization of the defense method, as follows:

Tab. R4: ASR of TRAIL on Additional Models.

ModelMethodFigStep-AudioAdvBench-AudioHarmBench
Qwen2.5-Omniw/o TRAIL40.83.817.5
w/ TRAIL32.40.912.6
Llama-Omniw/o TRAIL73.632.635.9
w/ TRAIL54.415.926.5

We did not adjust any hyperparameters for these additional models and directly applied the original settings to Qwen2.5-Omni and Llama-Omni. As shown in Tab. R4, TRAIL reduces ASR on both additional models across all three benchmarks. These results are reported in Appendix D.2.

Q5: Selection of k and Basis Stability.

A5: Thank you for your question. The parameter kk controls the number of basis vectors used to characterize the over-refusal subspace. A value that is too small may fail to capture sufficient safety-related information, whereas a value that is too large may introduce less relevant components, weakening the safety disentanglement effect. To investigate the impact of kk, we have supplemented the BRR results under different values of kk on the FigStep-Audio safety evaluation set, as shown below.

Tab. R5: BRR (%) under Different Values of kk on FigStep-Audio.

Metrick=1k=2k=3k=4
BRR (%)30272526

The cross-sample stability of the basis vectors is illustrated in Fig. 2(e–f) on page 4. As shown, the extracted basis vectors exhibit consistent patterns across different inputs while distinguishing harmful queries from benign ones. We have provided an analysis of this observation in Lines 230–232 of the original manuscript and will further elaborate on it in Section 3.2 of the revised paper.

AuthorsOpenReview ↗

Official Comment By Authors

Dear reviewer Se2k ,

Thank you for your valuable comments and constructive suggestions.

As the discussion deadline is approaching, we would appreciate your participation in the discussion so that we have sufficient time to address any further questions or concerns you may have.

If you feel that our response has addressed your concerns, would you kindly consider raising your rating score?

Thank you for your time and consideration, and we look forward to hearing from you.

Best Regards,

Authors of Submission 35924

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.