acceptodds
Withdrawn from ICLR 2027

ReasonBreak: Probing Vulnerabilities in Reasoning-Enabled Vision-Language-Action Models for Autonomous Driving

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-Language-Action (VLA) models with integrated reasoning have been proposed for end-to-end autonomous driving, assuming a tight coupling between reasoning and trajectory generation. However, the robustness of such systems under input perturbations remains largely unexplored. We show that these models are highly vulnerable to input perturbations, achieving up to 100% attack success rate (ASR) on reasoning in open-loop and up to 72% on trajectory manipulation in closed-loop simulation, leading to increased collision rates and degraded safety metrics. Using NVIDIA's recent Alpamayo models as representative industry-developed VLAs, we conduct a systematic black-box study of reasoning-enabled driving VLAs and evaluate how perturbations affect reasoning and driving behavior. We introduce ReasonBreak, a reasoning-aware evaluation framework that captures semantic and structural changes in reasoning, trajectory degradation, reasoning–trajectory coupling, and downstream safety impact. Our analysis reveals that reasoning and trajectory changes are only weakly coupled under perturbation, with substantial changes in one output surface often occurring without corresponding changes in the other. We also introduce a benchmark for evaluating attacks and defenses on reasoning–trajectory interactions in autonomous driving. Our results highlight the need for rigorous evaluation and improved defenses to ensure the safety of reasoning-enabled VLA systems in autonomous driving. Implementation code and data are available.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Vision-Language-Action models, Reasoning, Evaluation Protocol, Security, Safety

Reviews

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

This paper studies the robustness of reasoning-enabled driving VLAs. The authors propose ReasonBreak, an evaluation framework that treats reasoning and trajectory as two output surfaces. Reasoning is decomposed by an LLM into object, relation, implication, and planning fields to measure semantic deviation, alongside structural failures, including slowdown and DoS. Besides, trajectory degradation, reasoning–trajectory coupling, and downstream safety are evaluated separately. Using Best-of-N surface-form perturbations in open-loop evaluation and per-step random perturbations in closed-loop simulation, the paper reports high attack success rates, weak reasoning–trajectory coupling, and degraded safety metrics.

Critical strengths
  1. Reasoning-enabled driving VLAs are an important direction in industry. The paper evaluates reasoning and trajectory as two separate but interacting output surfaces, which is practically relevant for real-time driving. The finding of weak reasoning–trajectory coupling also offers useful insight into the faithfulness of reasoning in driving VLAs.
  2. The paper covers both open-loop and closed-loop evaluation, provides sensitivity analyses on the query budget, several thresholds, window length, perturbation strength, and input length, and validates the semantic evaluator with multiple LLM judges and a human study. The paper also honestly reports negative and mixed results and provides a simple mitigation baseline.
Critical weaknesses
  1. Improved Readability: [1] The format of all references in the main paper is inappropriate, except for Sec 4.2 Open-Loop (line 312). Might use \citep{} instead of \cite{}. [2] In Sec 2.1, a potential typo is meaningless '/' in 'black-box/visual attacks' (line 140). [3] In Sec 3.2.2 Open-loop, for definition of ASR (line 236), ∃\exists is only limited to ii, so how to decide ss which has two conditions? And in Sec 3.2.1 Reasoning Evaluation, cc has been noted as 'reasoning target' (line 209), but here cc is defined as target objective (line 238), so are these two cc the same meaning, and what is the relation between ss and cc if ss is Traj\text{Traj} because there is no definition of cc for trajectory evaluation beforehand in Sec 3.2.1. And the use of subscripts and superscripts for C\mathcal{C} is highly inconsistent, that is, the definitions indicated by the indices for C\mathcal{C} differ across Sec 3.2.1 Reasoning Evaluation/Trajectory Evaluation and Open-loop here. [4] In Sec 3.2.2 Closed-loop, there is a definition of 'excess deviation' (line 242), so is it different from perturbation-induced excess trajectory error (line 222) in Sec 3.2.1 Trajectory Evaluation, whose meaning is similar? And here is a conflict that oto_t denotes either reasoning rtr_t or trajectory output τt\tau_t (line 245), but oo has been noted as the trajectory output or ego state (line 221) in Sec 3.2.1 Trajectory Evaluation. [5] In Sec 4.2, (ri,τi)(r_i, \tau_i) and (rk,τk)(r_k, \tau_k) are used to represent perturbed input (line 320 and line 326), however, (r,τ)(r,\tau) are noted as benign input (line 175) in Sec 3.1, so should the variants of (r′,τ′)(r',\tau') be used here? [6] In Sec 5.2, the referred Table 3, Table 4, Figure 3, and Figure 4 are after the 9th page, which is outside the main paper scope, but it seems like that they do not belong to any section in Appendix. [7] In Appendix C.1, the decoding configurations are outside the paper column (line 1412 and line 1420), and in Appendix C.4, the closed vocabulary is under the same condition (line 1527).
  2. More Illustration: [1] In Sec 3.2.2 Closed-loop, there is no detailed information for the function Dcl(s,c)D_{\text{cl}}^{(s,c)}, that is, the concrete caculation process. [2] In Sec 4, there is no detailed information on perturbation injection, except for 'capitalization changes, word scrambling, and character-level noise' (line 297). There should be more illustration of how to achieve the perturbed textual input, what these perturbed texts look like, and how these perturbations might influence each type of reasoning surface and trajectory surface, such as slowdown. [3] In Sec 5.2, there is only a brief conclusion about the weak correlation between reasoning and trajectory. It would be better to have an analysis and explanation of this phenomenon, as it has been noted 'contrasts with the intended role of reasoning in reasoning-enabled VLA systems' (line 416). [4] There is no sensitivity analysis for the semantic-deviation threshold δsem\delta_{\text{sem}}, a related experiment is needed to illustrate this parameter setting. For the analysis of window size in Appendix B.4, the temporal consistency requirement kk is introduced here (line 1385), but in Sec 3.2.2 Closed-loop, the window size is noted as ww, so please illustrate the precise definition of 'window size' or the relation between kk and ww.
Questions for authors

Detailed in Weakness.

Reviewer_vAA5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

ReasonBreak evaluates textual perturbations against reasoning-enabled autonomous-driving VLAs, measuring semantic and structural reasoning changes, trajectory degradation, and downstream safety. Experiments cover Alpamayo1, Alpamayo1.5, and AutoVLA, including closed-loop simulation for the Alpamayo models. The paper reports substantial vulnerability and weak association between reasoning changes and trajectory-error changes.

Critical strengths
  1. The evaluation separates reasoning manipulation, trajectory degradation, and actual safety outcomes. This provides a useful diagnostic perspective and avoids treating every change in generated reasoning as a driving failure.
  2. Closed-loop evaluation provides evidence beyond textual instability. Single-query perturbations introduce additional safety incidents for Alpamayo1, demonstrating that some corruptions have consequential effects on driving. Evaluator checks and sensitivity analyses further support the breadth of the investigation.
Critical weaknesses
  1. How much of the open-loop ASR exceeds ordinary generation variability? Alpamayo evaluation uses temperature 0.6 and searches up to 100 perturbed executions against a benign reference output. A matched repeated-clean control is not reported. Consequently, semantic changes, reasoning-length increases, and trajectory-error increases may partly reflect stochastic generation selected through repeated querying. This affects the attribution and magnitude of the open-loop results, although the deterministic closed-loop experiments provide separate evidence of vulnerability.
  2. Does the DoS criterion distinguish missing reasoning from unavailable driving output? Section 3.2.1 checks for zero reasoning tokens, whereas the introduction describes failure to generate any tokens. This distinction is particularly important for AutoVLA, which supports direct action generation (https://arxiv.org/abs/2506.13757v1). The reported 100% DoS rate does not establish driving-output unavailability unless valid action tokens or trajectories are also absent; Table 4’s N/A entry does not identify which condition occurs.
Questions for authors
  1. What ASR results from repeated unperturbed generation under the same decoding configuration, query budget, and success criteria?
  2. Does the weak association persist when trajectory change measures the difference between benign and perturbed behavior rather than the change in ground-truth error?
  3. In AutoVLA’s DoS cases, are valid action outputs unavailable, or is explicit reasoning omitted? How is normal fast-thinking behavior distinguished?
  4. Why do the overall closed-loop semantic ASRs differ between Table 2 (56%/40%) and Table 20 at the stated default (k=3) (86%/78%)? Do the reference or success definitions differ?
  5. Which deployment boundary permits modification of the system template, and do the main closed-loop findings hold when trusted templates are fixed and only external textual inputs are editable?
  6. How does closed-loop reasoning evaluation distinguish direct reasoning corruption from appropriate responses to different observations after the rollouts diverge?
Reviewer_8PGQOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents ReasonBreak, a framework for testing vulnerabilities in reasoning-enabled vision-language-action (VLA) models for autonomous driving. It examines how meaning-preserving text changes affect models’ reasoning, predicted trajectories, and safety. The authors test Alpamayo1, Alpamayo1.5, and AutoVLA, using Best-of-N searches in open-loop experiments and single random perturbations in closed-loop simulations. Attacks achieve success rates as high as 100% for reasoning changes in open-loop tests and 72% for trajectory changes in closed-loop tests. Changes in reasoning correlate only weakly with trajectory deviations, suggesting the two outputs may be loosely connected under perturbation.

Critical strengths
  • The paper addresses an underexplored security challenge in reasoning-enabled driving VLAs. Its separate evaluation of reasoning, trajectory behavior, their interaction, and safety provides a useful framework. Defining trajectory attack success relative to benign behavior is also a thoughtful choice.
  • The closed-loop evaluation is especially valuable: it uses one perturbation per step, deterministic decoding, and temporal consistency requirements. Even without searching for attacks at every step, trajectory ASR reaches 72% for Alpamayo1 and 48% for Alpamayo1.5, with additional safety violations in simulation.
  • The paper’s ablations examine input modalities, attack strategies, perturbation strength, evaluator reliability, and a simple input-normalization defense. Together, they provide a broader picture of the vulnerabilities.
Critical weaknesses
  • The open-loop results lack an essential baseline. Because inference uses stochastic decoding (temperature 0.6), and attack success is defined as any of up to 100 perturbed queries differing sufficiently from a benign output, it is unclear how much of the reported ASR comes from perturbations rather than normal sampling. The semantic-deviation threshold is also unspecified. Although the authors validate their LLM evaluator, the human study is small, covering only 10 reasoning pairs and five decomposition examples.
  • The practical threat model needs clarification. The experiments perturb system and user prompts (Table 21), rather than realistic attacker-controlled inputs such as speech transcripts or navigation commands. It is unclear whether attackers could access privileged instructions in deployment, and default-strength perturbations do not resemble typical transcription errors. Navigation commands are more realistic but receive limited attention. Comparisons with non-reasoning models would help determine whether these vulnerabilities are specific to reasoning-enabled VLAs or reflect a broader sensitivity to noisy text.
  • The safety conclusions overstate the evidence. While the paper emphasizes increased collision rates, most open-loop reasoning attacks lower the collision proxy for Alpamayo1 and AutoVLA (Tables 3–4). The clearest increases come from trajectory-targeted attacks. These open-loop metrics are geometric proxies, not actual collisions. Closed-loop results are more meaningful, but cover few incidents and focus mainly on successful attacks. Confidence intervals and results across all rollouts would strengthen the claims.
  • Evaluation details are inconsistent, and evidence for coupling is limited. At the default temporal window, Table 2 reports much lower closed-loop reasoning ASRs than Table 20 (e.g., 48% vs. 78% for Alpamayo1 object reasoning). Success is also defined differently in Section 3.2.2 and Appendix B.4. The coupling claim rests on correlations between LLM-judged reasoning deviations and minimum ADE across six trajectories, which may mask distribution shifts. Concentrated semantic scores may further weaken the correlation. Closed-loop results cover only two Alpamayo models, so generality remains uncertain.
Questions for authors
  • Can the authors report benign-to-benign semantic deviations under identical stochastic decoding, along with the resulting Best-of-N control ASR? What semantic threshold is used in Table 1?
  • Why do Tables 2 and 20 show substantially different reasoning ASRs at the default temporal window, and which success definition is used for the main results?
  • Can the authors provide confidence intervals for the main ASRs and safety outcomes, and report closed-loop incident rates across all 50 rollouts, not just successful attacks?
  • Which text inputs could an attacker realistically change in a deployed driving system? Can the authors test perturbations limited to those inputs, ideally using realistic transcription or navigation errors?
  • Would the weak correlations remain when using mean ADE, trajectory-distribution changes, or deviations from the benign predicted trajectory instead of min-ADE? A scatter plot of the underlying score distributions would also help.
  • Does input normalization affect performance on clean inputs, especially navigation commands and out-of-vocabulary place names?
  • Can the authors provide closed-loop results for AutoVLA, or limit their closed-loop conclusions to Alpamayo? Please also discuss how Alpamayo1.5’s relatively high benign off-road rate affects the safety comparisons.

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

Related papers

Discussion (0)

Sign in to comment.