Research on Multi-Agent Competition Based on Large Language Models

Large Language Models (LLMs), with their advanced capabilities in semantic understanding, dynamic strategy generation, and complex behavior simulation, represent a transformative tool for studying multi-agent interactions. Their exceptional performance in areas such as natural language processing and social behavior modeling enables the simulation of socio-economic phenomena through LLM-driven agents. However, existing research has primarily focused on agent collaboration, leaving a significant gap in the exploration of competition mechanisms�an essential driver of social evolution and economic innovation. To address this theoretical gap, this paper introduces a bidirectional competition-cooperation framework, where service agents and experience agents are treated as equal participants in a game. A three-dimensional competitive space of "resources-influence-information" is constructed. Leveraging LLM-driven strategy generators and reinforcement learning optimization mechanisms, service agents can dynamically adjust strategies such as pricing and menus, while experience agents influence market dynamics through behaviors such as social network dissemination and hidden information exploration. This research overcomes the unidirectional constraints of traditional models, highlighting the symbiotic evolution of competition-cooperation behaviors in digital societies. It also provides a computationally viable theoretical tool for applications such as platform economy governance and metaverse business simulation.