Research

This study presents a novel traffic modeling framework that combines Large Language Models (LLMs) with real-world trajectory data to reproduce human-like driving behavior realistically. Using a hybrid driver clustering method (SOM + K-means++) and fine-tuned GPT-4.1 nano, the model generates trajectories based on contextual reasoning via Shared Memory and Chain-of-Thought structures. Compared to the conventional IDM model in SUMO, the LLM-based model exhibits superior accuracy (in terms of speed, acceleration, and headway), enhanced safety (with fewer critical TTC events), and improved interpretability through natural-language decision explanations. The findings highlight the potential of LLMs for behavioral traffic simulation and autonomous vehicle applications.