An LLM-assisted real-time simulation framework for demand-responsive transport operations

Demand-responsive transport (DRT) is a dynamic service whose performance depends not only on dispatching and routing algorithms but also on how externally submitted demand is captured, interpreted, and incorporated into operations. Although prior studies have largely emphasized algorithm-level optimization using pre-defined demand scenarios, integrated validation of externally submitted demand, user- and operator-facing interfaces, and operational analytics remains limited. This study proposes an LLM-assisted real-time DRT operation framework comprising Service User, System Manager, and Real-time DRT Simulator layers. The Service User layer uses function calling to transform unstructured trip requests into structured dispatch parameters. The System Manager combines function-calling-based quantitative analysis with retrieval-augmented generation to support natural-language access to operational data and domain knowledge. Based on Discrete Event System Specification, the simulator processes externally submitted and simulation-generated demands within a unified event structure under common dispatching and routing logic. Three experiments were conducted. Experiments 1 and 2 functionally evaluated the rider request-processing interface and the function-calling-based quantitative-query path of the System Manager, yielding average response latencies of 1.22 and 2.21 s, respectively. Experiment 3 independently evaluated the simulator through a system-level stress test using context-based synchronized demand. Across all scenarios, synchronized demand increased waiting times for the target passenger group. Although this group represented approximately 20% of successfully served passengers, it accounted for 26.5–28.5% of total waiting time. The effect appeared largely localized and was partially obscured in aggregate metrics because background-passenger performance was stable. These results demonstrate the functional feasibility of the LLM-assisted interfaces under the prototype conditions and show that synchronized-demand stress testing can reveal localized delay burdens that aggregate evaluations may overlook.