Crash Root-Cause Identification via Trace-Rewarded Causation-Chain Reasoning Large Language Model
提出一种基于大语言模型和痕迹奖励函数的因果链推理方法，通过蒙特卡洛树搜索和策略优化，将碰撞根因识别准确率从0.427提升至0.870，为交通安全管理提供支持。
Road traffic crash is one of the top leading causes of death worldwide. To support the deployment of modern safety improvements such as Highly Automated Driving (HAD), collision analysis requires to reveal crash formation and find the root-causes. This study explores Large Language Model (LLM)-based techniques to reconstruct the crash causation chain, and reveal the root-causes underlying crash occurrence. By designing trace-reward functions, a domain reasoning model base upon DeepSeek-R1-Distill-Qwen-1.5B is constructed to identify the root-causes. Specifically, the trace-rewards are constructed from accuracy in recognizing crash types and entities, accuracy in extracting crash-related behaviors, and alignment degree of behavior and root-cause. Monte Carlo Tree Search (MCTS) is then employed to broaden the exploration of potential root-causes and build inference paths. Finally, Group Relative Policy Optimization (GRPO) is applied to identify the optimal inference traces through training the model. Empirical analyses are conducted using Multi-Modal Accident Video Understanding (MM-AU) dataset. The results show that the proposed method raises the Micro Accuracy of root-cause identification from 0.427 to 0.870 and improves the Macro Recall from 0.389 to 0.852. This demonstrates that the proposed method effectively enhances the benchmark model's ability to understand the process of crash formation. Based on the identified root-causes, further analyses and discussions are conducted, providing effective support for traffic safety management and the development of preventive strategies.