Computer generated force (CGF) is the core component of military simulation systems. Traditional modeling methods suffer from bottlenecks including rigid knowledge representation, scarcity of high-quality samples, insufficient modeling of decision complexity, and lack of behavioral evolution capability. Large language models provide a new paradigm to address these issues. This paper systematically clarifies the enabling paths of large models from three dimensions: data and knowledge enhancement, decision intelligence generation, and capability iterative evolution. Focusing on five key modules: perception, decision-making, action, role, and memory, it elaborates on the large language model based CGF decision-making behavior modeling framework, sorts out the technical implementation routes and representative research achievements of each module, and summarizes key technical characteristics and application status. Furthermore, it proposes future research directions from five aspects: decision real-time performance, decision quality, decision fidelity, decision evaluation system, and decision risk control, which can provide a systematic reference for intelligent CGF research and the intelligent upgrading of military simulation.