Teaching self-driving cars to talk without slowing down

Self-driving cars that coordinate with each other typically either share raw perception data (bulky) or exchange language (slow, lossy). This work identifies a specific failure mode—agent identity confusion when latent representations blend across vehicles—and proposes LACO, a training-free approach that adapts existing driving models to collaborate. Using three techniques (iterative deliberation, selective information transmission, and knowledge distillation), LACO cuts communication and inference latency on CARLA while maintaining safety and efficiency gains.