Crash Risk Prediction Guided by Large Language Models Using Pre-Crash Trajectories

This study develops a crash risk prediction model at the individual vehicle level by leveraging the reasoning capability of large language models (LLMs). Instead of using the LLM itself for online prediction, the proposed framework converts LLM reasoning into structured supervision and distills it into a lightweight temporal graph network, which is then calibrated with real crash data. The framework is evaluated using pre-crash vehicle trajectories from 109 real-world crashes captured in multi-year drone videos at freeway merging and weaving segments in Nanjing, China. Results show that the proposed model outperforms representative benchmarks, including direct LLM inference, direct supervised training on real crash data, and traditional surrogate safety measures. Compared with direct supervised training using the same student architecture, it increases AUC by 7.8% and reduces MAE by 28.0%. Moreover, the proposed model is particularly effective when real crash data are limited. When only 10% of the fine-tuning data is used, it achieves the best performance among the compared methods, reducing MAE by 19.6% relative to the LLM baseline and by 38.0% relative to direct supervised training. These findings suggest that the proposed framework can improve prediction performance and data efficiency when real crash data are limited.