What question did this study set out to answer?

The review aims to understand how large language models are applied in intelligent transportation systems to enhance decision-making and coordination.

June 19, 2026

TrafficMind: A System-Oriented Review of Large Language Models for Intelligent Transportation Systems

Puntos clave

The review aims to understand how large language models are applied in intelligent transportation systems to enhance decision-making and coordination.
Conducted a systematic review of applications of large language models in intelligent transportation systems.
Analyzed how LLMs contribute to perception, forecasting, control, planning, and coordination.
Developed the TrafficMind framework for system-level integration of LLMs in transportation.
LLMs mainly provide coherence among heterogeneous data and human inputs, rather than focusing on numerical optimization.
In safety-critical applications, such as traffic signal control, LLMs are more effective when used for high-level decision support rather than in closed-loop control.
Emerging designs involve combining LLMs with reinforcement learning and knowledge graphs for better traffic management.

Resumen

Large language models (LLMs) are beginning to appear across a wide range of intelligent transportation system (ITS) applications, from traffic monitoring and prediction to signal control, planning, and human–machine interaction. Unlike conventional learning-based components that are designed for narrowly defined tasks, LLMs are increasingly used to organize information, interpret rules and intentions, and support decision-making across multiple parts of a traffic system. This paper reviews recent work on LLM-enabled ITS and uses these studies to develop a system-level perspective, referred to as TrafficMind, in which language models function as a shared semantic and reasoning layer. By examining how LLMs are currently applied in perception, forecasting, control, planning, and coordination, we find a consistent pattern: their main contribution is not numerical optimization or real-time actuation, but the ability to connect heterogeneous data, regulatory constraints, and human inputs in a coherent and interpretable way. In safety-critical settings, including traffic signal control and urban air mobility (UAM), existing evidence also suggests that LLMs are most reliable when they are kept outside closed-loop control and instead used to support coordination, explanation, and high-level decision processes. We further review emerging system designs that combine LLMs with reinforcement learning (RL), knowledge graphs, and digital-twin platforms, allowing traffic decisions to be both data-driven and grounded in physical and regulatory structure. Beyond technical performance, we explore how these developments raise new questions for governance, certification, and the ethical deployment of models, particularly in systems where model outputs can impact real-world traffic outcomes. Overall, the TrafficMind perspective offers a way to interpret current progress and to frame how LLMs can be integrated into future ITS in a manner that is transparent, accountable, and compatible with real operational practice.

Preguntar a la IA

Me gusta

Guardar