AITP: Traffic Accident Responsibility Allocation via Multimodal Large Language Models

Researchers Zijin Zhou and Songan Zhang have unveiled AITP (Artificial Intelligence Traffic Police), a multimodal large language model designed to determine fault in traffic accidents. Unlike prior systems that merely describe or detect accidents, AITP performs multi-step causal reasoning grounded in traffic regulations. It enhances this capability through a Multimodal Chain-of-Thought (MCoT) mechanism, which breaks down complex accident scenarios into logical reasoning steps, and integrates legal knowledge via Retrieval-Augmented Generation (RAG), allowing it to pull relevant traffic laws during analysis.

To train and evaluate AITP, the team also created DecaTARA, a decathlon-style benchmark unifying ten interrelated traffic accident reasoning tasks. This dataset includes 67,941 annotated videos and 195,821 question-answer pairs, covering everything from basic detection to nuanced responsibility allocation. In extensive experiments, AITP achieved state-of-the-art performance across all three task categories—responsibility allocation, traffic accident detection (TAD), and traffic accident understanding (TAU). The paper was accepted at CVPR 2026 and is available on arXiv.