A Reasoning Layer for Traffic Operators: Leveraging Large Language Models to Predict, Detect, and Estimate Traffic States

Confirmation seminar presented by Zetong Zhang

The study investigates the potential of large language models (LLMs) in traffic monitoring, a domain built on three parallel tasks: predicting near-term conditions, detecting anomalies, and estimating states at unsensed locations. Integration with established numerical engines is examined at three levels: interpretation of outputs, orchestration within analytic pipelines through retrieval and checks, and direct inference. Under a unified evaluation–improvement framework, experiments are designed across prediction, anomaly detection, and estimation to evaluate how LLMs process traffic data, interact with numerical models, and generate operator-ready narratives while addressing challenges of context length, evidence integration, and robustness.