arXiv:2606.08684v1 Announce Type: new 
Abstract: We present BLUE, a minimal method for better language use in vision-language-action (VLA) models for autonomous driving (AD). Through extensive analysis, we reveal that language matters on only a small fraction of routes, but on those routes it can greatly improve or degrade performance. Generating language at every frame is therefore inefficient, since most computation is spent on frames that do not benefit from language. We further show that pretrained VLA hidden states potentially already encode whether language will benefit a given frame, even though scene complexity and kinematic features alone struggle to predict this. Based on this finding, BLUE trains a lightweight gate on frozen VLA hidden states to decide per frame whether to activate language generation or predict actions directly, without modifying the backbone or requiring additional human annotation. With just a 0.11M-parameter gate, BLUE sets a new state of the art on both benchmarks, achieving 76.2% success rate on Bench2Drive and 36 driving score on Longest6 v2, while delivering 2.54x inference speedup and 8.9% success rate improvement over the backbone. BLUE provides a practical path toward efficient language-augmented AD, showing that VLA models can retain the benefits of language at a fraction of the cost. Our code, data, logs and checkpoints are fully available on https://github.com/George-Ling3/BLUE.

تم تقديم BLUE، وهي طريقة بسيطة لتحسين استخدام اللغة في نماذج العمل بين اللغة والرؤية (VLA) للقيادة الذاتية، مما يكشف أن اللغة تؤثر بشكل كبير على الأداء في بعض الطرق. تستخدم هذه الطريقة بوابة خفيفة لتحديد متى يجب تفعيل توليد اللغة، مما يحسن الكفاءة الحسابية دون تعديل الهيكل الأساسي للنموذج.

La introducción de BLUE, un método mínimo para mejorar el uso del lenguaje en modelos de acción lenguaje-visión (VLA) para la conducción autónoma, revela que el lenguaje impacta significativamente en el rendimiento en rutas seleccionadas. Este método emplea una puerta ligera para determinar cuándo activar la generación de lenguaje, optimizando la eficiencia computacional sin alterar la estructura del modelo.

L'introduction de BLUE, une méthode minimale visant à améliorer l'utilisation du langage dans les modèles d'action langage-vision (VLA) pour la conduite autonome, révèle que le langage a un impact significatif sur la performance sur certaines routes. Cette méthode utilise une porte légère pour déterminer quand activer la génération de langage, optimisant ainsi l'efficacité computationnelle sans modifier l'architecture du modèle.

The introduction of BLUE, a minimal method for enhancing language use in vision-language-action (VLA) models for autonomous driving, reveals that language significantly impacts performance on select routes. This method employs a lightweight gate to determine when to activate language generation, optimizing computational efficiency without altering the model's backbone.