To address the deficiencies in posture prediction and decision-making capabilities of large language models (LLMs) in complex battlefield environments, this study takes the human mental simulation theory as its core foundation and proposes a combat decision-making method for LLMs that integrates mental simulation mechanisms with counterfactual reflection. The approach employs mental simulation-driven posture prediction, which combines full-parameter fine-tuning with military common sense, weapon parameters, and situational deduction pathways, enabling the LLM to generate high-confidence future posture predictions based on the current battlefield status. For decision optimization through counterfactual reflection, the method simulates "unexecuted" counterfactual scenarios for alternative decisions, compares causal differences between "executed" and "unexecuted" outcomes, and generates comprehensive decisions that balance "risk-reward" trade-offs. Simulation experiments in amphibious landing scenarios demonstrate that this method significantly enhances the posture prediction capability and combat decision-making performance of LLMs. By integrating mental simulation theory with counterfactual reflection, it effectively improves the decision-making efficacy of LLMs in complex battlefield environments, offering a new pathway for intelligent command and decision-making systems.