arXiv:2605.29114v1 Announce Type: cross 
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 realistic input perturbations remains largely unexplored. We show that these models are highly vulnerable to realistic input perturbations, achieving up to 89% attack success rate (ASR) on reasoning 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 the first systematic black-box study of reasoning-enabled VLA models under realistic textual input corruptions, evaluating their impact on reasoning and driving behavior. We introduce a reasoning-aware evaluation framework capturing both semantic and structural aspects of reasoning, along with safety-centric measures. 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.

تكشف دراسة حديثة بعنوان 'ReasonBreak' عن وجود ثغرات كبيرة في نماذج العمل بلغة الرؤية (VLA) المستخدمة في القيادة الذاتية، مما يبرز حساسيتها للت perturbations المدخلة الواقعية. تشير الأبحاث إلى أن هذه النماذج يمكن أن تتعرض لمعدل نجاح هجوم يصل إلى 89% في مهام التفكير و72% في التلاعب بالمسار، مما يؤدي إلى زيادة معدلات الاصطدام وتدهور مقاييس السلامة.

Un estudio reciente titulado 'ReasonBreak' revela vulnerabilidades significativas en los modelos de Acción-Lenguaje-Visión (VLA) utilizados para la conducción autónoma, destacando su susceptibilidad a perturbaciones de entrada realistas. La investigación indica que estos modelos pueden experimentar hasta un 89% de tasa de éxito en ataques sobre tareas de razonamiento y un 72% en la manipulación de trayectorias, lo que lleva a un aumento en las tasas de colisión y métricas de seguridad comprometidas.

Une étude récente intitulée 'ReasonBreak' révèle des vulnérabilités significatives dans les modèles Vision-Language-Action (VLA) utilisés pour la conduite autonome, mettant en lumière leur sensibilité aux perturbations d'entrée réalistes. La recherche indique que ces modèles peuvent subir jusqu'à 89 % de taux de réussite d'attaque sur les tâches de raisonnement et 72 % sur la manipulation de trajectoire, entraînant des taux de collision accrus et des métriques de sécurité compromises.

A recent study titled 'ReasonBreak' reveals significant vulnerabilities in Vision-Language-Action (VLA) models used for autonomous driving, highlighting their susceptibility to realistic input perturbations. The research indicates that these models can experience up to an 89% attack success rate on reasoning tasks and a 72% rate on trajectory manipulation, leading to increased collision rates and compromised safety metrics.