LLM-Raft: Enhancing Urban Traffic Efficiency and Safety through Decentralized Coordination of Autonomous Vehicles

Urban areas face persistent challenges of traffic congestion and safety, which hinder efficiency and quality of life. Coordinated autonomous vehicles (AVs) offer a promising solution, but achieving robust, decentralized coordination in dynamic urban settings remains a significant hurdle. This paper introduces LLM-Raft, a novel framework designed to enhance urban mobility by enabling LLM-powered AVs to coordinate their actions safely and efficiently. Inspired by the Raft algorithm, LLM-Raft allows vehicles to generate and agree upon ``traffic narratives''---human-like, structured propositions of their intent and justification. This semantic consensus mechanism allows for more intelligent and predictable group behaviors without a central coordinator. We validate our framework in realistic urban traffic simulations. The results show that LLM-Raft improves key urban mobility metrics, reducing collision rates by 40-50\% and task completion time by 20-30\% compared to uncoordinated baselines. Our work presents a viable path toward more collaborative and resilient multi-agent systems, contributing to the development of safer and more efficient urban transportation networks. Code is available at https://github.com/shxingch/llm-raft.