LLM-TrafficBrain: An Information-Centric Framework for Dynamic Signal Control with Large Language Models

Dynamic and context-aware traffic signal control remains a significant challenge in intelligent transportation systems (ITS), particularly under rapidly evolving traffic patterns and unexpected events. This study proposes a novel framework integrating Large Language Models (LLMs) with real-time traffic sensing to enable semantic traffic signal scheduling. By translating structured traffic state data—including queue lengths, temporal context, and special events—into natural language prompts, the LLM functions as a reasoning agent to generate adaptive signal control policies. The framework operates within a closed-loop feedback system, facilitating real-time adjustments based on dynamic traffic conditions. Validation through simulation-based case studies demonstrates that the proposed approach achieves competitive or superior performance compared to conventional rule-based and reinforcement learning methods, measured by average delay reduction and throughput improvement. Additionally, it offers enhanced interpretability (via natural-language decision logs) and operational flexibility (e.g. handling priority requests for emergency vehicles). This work highlights the potential of LLMs as semantic planners for urban traffic control and contributes a scalable, prompt-driven architecture for intelligent intersection management.