arXiv:2511.20720v1 Announce Type: cross 
Abstract: Vision-Language Action (VLA) models unify perception, reasoning, and trajectory generation for autonomous driving, but suffer from significant inference latency due to deep transformer stacks. We present DeeAD, a training-free, action-guided early-exit framework that accelerates VLA planning by evaluating the physical feasibility of intermediate trajectories. Instead of relying on confidence scores, DeeAD terminates inference when predicted trajectories align with lightweight planning priors (e.g., Navigation or Low-precision Planning) within a tolerable deviation (<2m). To improve efficiency, we introduce a multi-hop controller that adaptively skips redundant layers based on the change rate of scores. DeeAD integrates into existing VLA models, such as ORION, without requiring retraining. Experiments on the Bench2Drive benchmark demonstrate up to 28% transformer-layer sparsity and 29% latency reduction, while preserving planning quality and safety.

تم تقديم DeeAD كإطار للخروج المبكر الديناميكي لنماذج العمل بلغة الرؤية (VLA)، بهدف تحسين كفاءة القيادة الذاتية من خلال تقليل زمن الاستدلال. يقوم هذا الإطار بتقييم الجدوى الفيزيائية للمسارات الوسيطة، مما يسمح بإنهاء الاستدلال مبكرًا عندما تتماشى مع أولويات التخطيط، مما يحقق تخفيضات كبيرة في زمن الاستدلال واستخدام طبقات المحولات.

DeeAD se ha presentado como un marco de salida temprana dinámica para modelos de Acción Lenguaje-Visión (VLA), con el objetivo de mejorar la eficiencia de la conducción autónoma al reducir la latencia de inferencia. Este marco evalúa la viabilidad física de las trayectorias intermedias, permitiendo la terminación anticipada de la inferencia cuando se alinea con prioridades de planificación, logrando reducciones significativas en la latencia y el uso de capas de transformadores.

DeeAD a été introduit comme un cadre de sortie précoce dynamique pour les modèles d'Action Vision-Langage (VLA), visant à améliorer l'efficacité de la conduite autonome en réduisant la latence d'inférence. Ce cadre évalue la faisabilité physique des trajectoires intermédiaires, permettant une terminaison précoce de l'inférence lorsqu'elle est alignée avec des priorités de planification, atteignant des réductions significatives de latence et d'utilisation des couches de transformateurs.

DeeAD has been introduced as a dynamic early exit framework for Vision-Language Action (VLA) models, aimed at enhancing the efficiency of autonomous driving by reducing inference latency. This framework evaluates the physical feasibility of intermediate trajectories, allowing for early termination of inference when aligned with planning priors, achieving significant reductions in latency and transformer-layer usage.

DeeAD: Dynamic Early Exit of Vision-Language Action for Efficient Autonomous Driving

One More Thing in AI – Your Shortcut to AI Mastery

DeeAD: Dynamic Early Exit of Vision-Language Action for Efficient Autonomous Driving

Was this article worth reading? Share it

One More Thing in AI

LucidQuery AI

Dyad

LangWatch

The Visualizer

Attentive AI

Ready to build your own newsroom?