LLM-Raft: Enhancing Urban Traffic Efficiency and Safety through Decentralized Coordination of Autonomous Vehicles
Abstract
Urban areas face persistent challenges of traffic congestion and safety, which hinder efficiency and quality of life. Coordinated autonomous vehicles offer a promising solution, but achieving robust, decentralized coordination in dynamic urban settings remains difficult.
LLM-Raft introduces a framework that enables LLM-powered autonomous vehicles to generate and agree on human-like traffic narratives that describe intent and justification. Inspired by the Raft algorithm, this semantic consensus mechanism improves predictability without relying on a central coordinator.
In realistic urban traffic simulations, the method reduces collision rates by 40-50% and shortens task completion time by 20-30% compared with uncoordinated baselines.
Overview
LLM-Raft enables autonomous vehicles to exchange human-readable traffic narratives that expose intent, anticipated motion, and justification instead of relying on opaque local policies alone.
The framework adapts Raft into a decentralized semantic-consensus protocol, allowing multiple vehicles to align on a shared interpretation of the scene before they act.
This coordination layer improves both safety and efficiency in dense urban environments, yielding fewer collisions and faster task completion in simulation.
BibTeX
@inproceedings{zhou2025llmraft,
author = {Lingfeng Zhou and Shuaixing Chen and Jin Gao and Dequan Wang},
title = {{LLM}-Raft: Enhancing Urban Traffic Efficiency and Safety through Decentralized Coordination of Autonomous Vehicles},
booktitle = {UrbanAI: Harnessing Artificial Intelligence for Smart Cities},
year = {2025}
}