LLM-DR: An On-Demand Delivery Simulation Framework for Studying Heterogeneous Rider Behavior and Collective Mobility with Persona-Grounded Large Language Model Agents

Abstract

On-demand delivery platforms are computational social systems in which dispatch rules and repeated rider actions jointly shape service and urban mobility. Riders choose whether to reject offers, accept one order, or bundle several orders; assignments then change their availability and other riders' opportunities. Modeling this feedback requires empirical rider heterogeneity, shared-order interaction, and road-network execution to be represented together. Delivery optimization and large language model (LLM) mobility generation provide useful foundations, but leave this combination insufficiently examined. We develop Large Language Models as Delivery Riders (LLM-DR), an agentic simulation framework for benchmarking rider policies under a common order stream and dispatch protocol. Persona-grounded LLM agents provide a reference implementation, using empirical profiles, local observations, short-term memory, and route-tool estimates to propose order actions. The environment resolves assignment conflicts, constructs pickup and drop-off sequences, and executes tool-supported routes. Analysis of 9,025 Beijing orders and a sample of 2,508 routed delivery waves identifies four rider personas. A controlled single-day reconstruction places 60 agents, equally distributed across personas, in an observed stream of 8,656 orders. Persona groups complete 16-30 orders per rider, with average bundle sizes of 1.116-2.212. Distributional comparisons retain several empirical persona orderings in order volume, bundling, and routed-wave distance. Fleet-level outcomes show displaced demand, service-occupancy, and completion peaks, alongside a shared central service core and differentiated spatial reach. These results connect data-derived rider strategies to executable action chains and collective mobility patterns within a shared computational social system.

arXiv comment17 pages, 7 figures, 2 tables. Code and de-identified data: https://github.com/Nicholas0027/LLM-DR