Hexagonal grid wargame simulations are an important tool for military training and strategic decision-making. This research uses the classic wargame "A Victory Lost" (AVL) as a carrier

focusing on its command sign extraction

control area

supply lines

and unique battle result tables

and constructs a reinforcement learning environment based on the Vassal engine. On this basis

a comparative experiment with the assistance of a large language model (LLM) for decision-making was designed and implemented. This research quantitatively analyzing the impact of LLM on win rate

score progress

and tactical turning points. Experimental results show that LLM assistance can significantly improve the win rate of the using side and change the timing of the turning points of the battle situation. This research also provides a reusable technical path for the development of AI in Vassal modules. It verifies the effectiveness and limitations of large language models in complex turn-based decision support and lays a foundation for subsequent parameter optimization and multi‑agent collaboration.