Dynamic LLM-Agent Network: An LLM-agent Collaboration Framework with Agent Team Optimization

Overview

This research paper proposes a novel framework called Dynamic LLM-Agent Network (DyLAN) for enhancing the collaborative capabilities of large language models (LLMs). DyLAN allows LLMs to work together on complex tasks by dynamically forming teams of agents and facilitating multi-round interactions. The key innovation lies in the dynamic architecture, which enables inference-time agent selection and an early-stopping mechanism, thereby optimizing performance and efficiency. DyLAN also introduces an unsupervised Agent Importance Score that automatically evaluates the contribution of each agent, enabling the selection of the most effective team composition. The authors demonstrate the effectiveness of DyLAN on various tasks, including arithmetic reasoning, general reasoning, and code generation, highlighting significant performance improvements compared to baseline methods. They also delve into the data efficiency and stability of the framework, showcasing DyLAN's robustness and potential for real-world applications.