Crash root-cause identification via trace-rewarded causation chain reasoning large language model

Area of Science:

• Artificial Intelligence
• Traffic Safety Engineering
• Computer Vision

Background:

• Road traffic crashes are a leading global cause of death.
• Modern safety systems, like Highly Automated Driving (HAD), require advanced collision analysis.
• Identifying crash root-causes is crucial for effective safety improvements.

Purpose of the Study:

• To explore Large Language Model (LLM)-based techniques for reconstructing crash causation chains.
• To identify the root-causes of road traffic crashes.
• To support the development of advanced traffic safety management and preventive strategies.

Main Methods:

• A domain reasoning model was constructed using DeepSeek-R1-Distill-Qwen-1.5B with designed trace-reward functions.
• Trace-rewards were based on accuracy in crash type/entity recognition, behavior extraction, and behavior-root-cause alignment.
• Monte Carlo Tree Search (MCTS) and Group Relative Policy Optimization (GRPO) were employed for root-cause exploration and optimal inference path identification.

Main Results:

• The proposed LLM-based method significantly improved Micro Accuracy of root-cause identification from 0.427 to 0.870.
• Macro Recall for root-cause identification was enhanced from 0.389 to 0.852.
• The study demonstrated an enhanced ability to understand crash formation processes.

Conclusions:

• The developed LLM-based approach effectively reconstructs crash causation chains and identifies root causes.
• This method provides robust support for traffic safety management and the creation of preventive strategies.
• The findings contribute to the advancement of safety analysis for systems like Highly Automated Driving (HAD).