Agentic Large Language Models for Decentralized Multi-agent Games

Language is a ubiquitous tool that is foundational to reasoning and collaboration, ranging from everyday interactions to sophisticated problem-solving tasks. Establishing a common language in multi-agent systems is an important foundation for facilitating desired coordination and strategies. In this work, we extend the capabilities of large language models (LLMs) by integrating them as powerful reasoning devices within multi-agent decision-making processes. We propose a systematic framework focused on key integration practices involving dynamic prompting techniques, multi-modal information processing, agentic roles and tool usage, and alignment methods specialized towards multi-agent objectives. We evaluate these design choices through extensive experimentation on classic game settings with key underlying social dilemmas and game-theoretic considerations. Our findings affirm the importance of the non-trivial design choices made by our proposed framework, which target alignment with specific solution concepts.