Design and Implementation of Game Decision-making System Based on Personified LLM Agents

As the core carrier of multi-agent body collaboration and competition, game decision systems have significant value in economics, artificial intelligence and complex system modeling. Traditional methods have significant limitations in terms of strategy diversity, dynamic adaptability and interpretability, which make it difficult to cope with the real-time evolution needs of complex game scenarios. This study propose a game decision-making system based on personified large language model (LLM), which realizes dynamic strategy aggregation through the synergistic mechanism of competitive, cooperative and analytical agents combined with a trainable meta-controller. The system adopts a layered architectural design and constructs a three-level interaction system containing environment perception, strategy generation and decision fusion, in which the personalized agents generate strategies through differentiated mathematical models and the meta-controller realizes context-aware weight allocation based on a layered attention mechanism. Experiments show that the system achieves an average score of 51.75 in a benchmark test containing 100 gaming scenarios, which is an improvement of 266.7% over the baseline model and exhibits significant advantages in dynamic strategy adaptation and decision interpretability. The core contribution of this study is to propose a scalable framework that integrates symbolic rules and neural networks, which provides a new technical idea for multi-agent body game research.