Accident causality chain mining: A collaborative multi-agent framework enhanced by knowledge-guided and counterfactual reasoning

Accident causal chain mining can accurately reconstruct the underlying logical event sequence leading to an accident, providing targeted evidence for accident prevention. However, when large language models (LLMs) are used for this task, their insufficient incorporation of domain knowledge and limitations in single-step reasoning often lead to an inaccurate identification of root causes and a tendency to oversimplify complex causal relationships as a single chain. In this paper, we propose a novel multi-agent collaborative framework with knowledge-guided and counterfactual reasoning (MAC-KCR). First, a knowledge-guided collaborative reasoning module is designed to enhance task-related information. This is achieved by leveraging theoretical knowledge of causal chains and facilitating interaction between two specialized agents: a cause analyst and a causal chain reasoner. Second, a counterfactual reasoning and logical verification module is introduced to refine the generated causal chain through multi-iterative feedback between another pair of agents: a counterfactual reasoner and a logical verifier. Experimental evaluations on electrical personal accident cases demonstrate the effectiveness of our framework in accident causal chain reasoning (direct cause identification: average F1 ≥ 0.92; full-chain mining: average F1 ≥ 0.75). We conclude that accurate causal chain mining critically depends on the collaborative, dynamic, and iterative error-correction mechanism inherent in the MAC-KCR framework, as a single module cannot overcome the challenges posed by long-chain dependencies and heterogeneous semantics.