Large language models (LLMs) demonstrate strong reasoning and planning capabilities in static textual contexts, yet they struggle significantly with dynamic decision-making tasks involving spatial elements, such as point selection in military simulations. These limitations arise from their reduced capacity to integrate real-time geographic data and adapt to spatial conditions, which can lead to crucial errors in positioning decisions. Such deficiencies may result in missed opportunities for tactical advantages, increased vulnerability, and diminished overall effectiveness in combat scenarios. To mitigate these issues, this paper presents the Geo-Commander framework, an innovative multi-task agent to combat simulations by integrate the ReAct reasoning mechanism and spatial encoding. The Geo-Choice module of this framework employs hexagonal grid encoding for preliminary location screening, enabling the agent to establish spatial constraints early in the decision-making process. The ReAct chain of this framework incorporates detailed geographic insights into the reasoning loop, yielding interpretable decisions for point selection. We validate the framework through experiments that reveal substantial performance improvements in both static point selections and real-time dynamic command tasks within a tank detachment combat simulation environment. Results indicate that Geo-Commander consistently surpasses control groups across various metrics, including selection quality, win rate, and overall combat effectiveness. These performance metrics highlight the framework’s potential to meet the demands of dynamic combat environments, ultimately confirming the feasibility of integrating spatial reasoning within LLM frameworks and opening avenues for advancements in multi-agent geospatial intelligence systems and battlefield decision-making support.
Chen et al. (Fri,) studied this question.