Leveraging large language models to enhance multi-agent risk assessment in supply chain networks

We propose a novel large language model (LLM) enhanced framework, MARS (Multi-Agent Risk assessment in Supply chain networks), for risk assessment and integration of both structured and unstructured factors for logistic hub site selection across a target territory, using the southeastern U.S. states as a testbed. While structured factors such as cost, distance, traffic accidents, traffic congestion, and crime rates can be directly computed, unstructured severe weather event narratives need to be interpreted semantically through LLMs. We introduce a multi-agent architecture featuring three specialised agents, RiskAgent, FeedbackAgent, and RevisionAgent, that collaborate through a feedback-revision loop to convert raw extreme weather event narratives into fine-grained risk severity levels. By integrating these severity levels with structured indicators via aggregation, the proposed method enables interpretable and risk-aware ranking of candidate hubs, thereby supporting informed decision-making for logistic hub site selection.