LLM agents for coordinated decision-making in cooperative-competitive games

The cooperative adversarial game environment is a key application scenario for multi-agent systems, requiring agents to collaborate with allies while adapting to dynamic opponent strategies. To address challenges such as partial observability, complex game dynamics, and high environmental uncertainty, this paper proposes a collaborative decision-making framework for large language model (LLM) agents in cooperative-competitive game scenarios. The framework introduces the following innovations: First, this paper designs a distributed multi-agent decision-making architecture to tackle partial observability challenges, enabling each agent to independently analyze and make decisions, thereby enhancing system flexibility and adaptability. Second, this paper develops a hybrid memory module to address dynamic environmental uncertainty, equipping LLM agents with dynamic learning and adaptive capabilities to improve decision-making performance in dynamic game scenarios. Third, to manage the complex game dynamics, this paper constructs a game knowledge base and multi-level reasoning mechanism, leveraging LLMs' natural language understanding and reasoning abilities to identify critical points and formulate proactive strategies. Experimental results demonstrate that this framework achieves efficient coordination and superior adversarial performance across several game scenarios, providing new insights for strategy design in complex multi-agent systems.