Project Summary: Rural and tribal roadways in the United States experience disproportionately high crash incidents due to a combination of infrastructure challenges, limited resources, and incomplete crash reporting. Traditional statistical, machine learning (ML), and deep learning (DL) models have struggled to address these issues because they rely heavily on structured data, perform poorly with incomplete or imbalanced datasets, and often fail to leverage the rich information contained in crash narratives. Large language models (LLMs) offer an alternative by reframing crash risk analysis as a text reasoning problem, enabling the extraction of contextual insights from narratives, the imputation of missing or inconsistent fields, and the integration of structured and unstructured data into unified predictive frameworks. This proposal aims to develop an LLM-based crash risk analysis system utilizing North Carolina’s statewide police-reported crash data as a foundation, with the goal of enhancing the accuracy, interpretability, and robustness of rural crash risk assessment. The project will proceed through four main tasks: crash data enhancement and model adaptation, predictive model development, model validation, and the design of an implementation plan for real-time crash risk warnings in connected vehicle (CV) environments.