SafeTraffic Copilot: adapting large language models for trustworthy traffic safety assessments and decision interventions
Yang Zhao, Pu Wang, Yibo Zhao, Hongru Du, Hao Yang
Abstract
Predicting expected traffic crashes and designing targeted interventions are highly challenging due to the inherent complexity of crash data and persistent concerns over the prediction trustworthiness. We introduce SafeTraffic Copilot that adapts Large Language Models (LLMs) to perform expected crash prediction as a text-reasoning task, then attribute critical features for targeted safety interventions. Within the Copilot, SafeTraffic LLM is customized then fine-tuned on the textualized SafeTraffic Event dataset, which consists of 66,205 real-world crash cases with 14.5 million words from five U.S. states. Across multiple prediction tasks including crash type, severity, and number of injuries, SafeTraffic LLM demonstrates a 33.3% to 45.8% improvement in average F1-score over existing works. To interpret these results and inform safety interventions, we introduce SafeTraffic Attribution, a sentence-level feature-attribution framework enabling conditional "what-if" risk analysis. Findings reveal that alcohol-impaired driving is the leading factor for severe crashes, with impairment-related and aggressive behaviors contributing nearly three times more risk than other behaviors. Furthermore, SafeTraffic Attribution identifies critical features during fine-tuning, guiding crash data collection strategies for continual improvement. SafeTraffic Copilot enables prediction and reasoning of conditional crash risks through foundation models, thereby supporting traffic safety improvements and offering clear advantages in generalization, adaptation, and trustworthiness.