TRAFFICGEN: a multi-agent LLM orchestration for smart mobility and emergency corridor pre-emption

Abstract

This research work introduces TrafficGen, a proof-of-concept software prototype that explores the possibility of adaptive traffic signal management using LLM-based agentic orchestration. The framework is evaluated in a single-intersection, deterministic simulation environment that uses text to represent the various intersections and can operate with Gemini 2.5 Flash and LangGraph to implement a “Council of Agents” control architecture. TrafficGen uses a semantic translation layer to generate human-readable chain-of-thought reasoning logs for each signal decision, enabling it to manage complex traffic patterns. In empirical stochastic evaluations, TrafficGen was tested using four custom stress scenarios and 30 independent runs, yielding a mean reduction of 46.9% ± 2.8% in Wait Burden Score across all scenarios and runs (from 6,370 points to 3,380 points per run, compared to standard fixed-time signal controllers). Additionally, 100% of the Green Corridor maintained a success rate for dynamic emergency vehicle pre-emption. Architecturally, the system achieved a 100% decision override latency of 1,556 ms and a Green Corridor success rate of 100% for life safety evaluations. The quantitative results show the effectiveness of zero-shot semantic LLM orchestration compared to other LLM controllers with heuristic rules. These results come from a controlled single intersection environment and may not guarantee similar performance when deploying traffic management in municipalities generally, but TrafficGen paves the way for traffic management to move beyond black-box numerical paradigms and into explainable decision-making. Finally, this paper is an initial step towards adaptive traffic management for smart urban mobility.