UrbanMUDA: an LLM Agent-based Site Selection Approach for Urban Military Unit Deployment

Urban Military Unit Deployment (UrbanMUD) has emerged as a critical task for extracting actionable knowledge from multi-source urban data to support diverse urban combat simulations. However, current practices remain highly dependent on manual effort, which limits their scalability and efficiency. In this paper, we propose UrbanMUDA, a unified agent framework powered by large language models (LLMs), for automated UrbanMUD. UrbanMUDA integrates a domain-informed instruction set, constructed via heterogeneity-aware and few-shot-based generation, to support two core tasks: geospatial reasoning (GR) and structured scenario extraction (SSE). Furthermore, a tool-augmented iterative refinement module is introduced to improve reasoning quality through multi-step interaction with external resources, without modifying the underlying LLMs. We conduct extensive experiments on real-world datasets from NYC and TW, across both GR and SSE tasks. UrbanMUDA achieves substantial accuracy improvements over zero-shot baselines, reaching up to 650% on GR and 159% on SSE. Consistent performance is observed across five different LLM backbones, demonstrating strong robustness and generalizability in both English and Chinese deployment scenarios. Although the approach incurs moderate increases in inference latency, the trade-off is justified by significant gains in deployment reasoning quality. These results validate the effectiveness of UrbanMUDA for scalable and multilingual military deployment automation.