LLM-Driven Multi-Agent Coordinated Control for Urban Rail Transit Disruption Response

Short section delays in metro corridors can spread quickly through transfer hubs and return to operations as longer dwell times, local crowding, and downstream delay. Existing disruption–recovery studies often optimize train regulation, station control, and passenger guidance in separate modules, which makes it difficult to assemble a linked response that is both operationally coherent and auditable. This study develops an LLM-driven multi-agent decision framework for abnormal response in urban rail transit. Role-specific agents generate train-, station-, transfer-, and information-side measures; a safety-review agent removes infeasible actions; and an arbitration agent produces a structured DecisionCard for simulation backfeeding. A corridor-level mesoscopic model is built for the Shanghai Metro Line 2 section between East Nanjing Road and Lujiazui, covering four key transfer hubs. The evaluation combines five online LLM cases with a broader 3-by-3 delay–demand matrix and transfer-capacity sensitivity tests. Relative to B4-SOP-Limited, Online-MA-Rolling reduces total delay by 3.5–48.8% and peak transfer queues by 2.0–90.3% in the five online cases. Across the nine delay–demand combinations, the multi-agent strategy reduces total delay by 58.9% on average and transfer-queue peaks by 32.6% on average compared with B4. When transfer service capacity is reduced by 25%, however, queue relief collapses and the delay advantage can reverse under high load. The main contribution of the framework is, therefore, not the use of LLMs alone, but the production of linked, reviewable, and simulation-testable response plans within a clearly bounded operating envelope.