From Transmitting Bits to Communicating Intelligence: Enhancing Communication Efficiency in LLM-Powered Multi-Agent Systems

Applying Large Language Models (LLMs) to domain-specific applications greatly  benefits real-world problem-solving, underscored as the leading breakthrough in MIT  Technology Review’s 10 Breakthrough Technologies of 2024. To further tackle more  complex tasks, LLM-based multi-agent systems are proposed to employ multiple  intelligent agents with specific roles and collaborate on target tasks. However, this  practice has led to significant deployment challenges, including multi-agent  organization, communication efficiency, distributed fine-tuning, and dynamic  environments.  This proposal is dedicated to addressing these multifaceted issues through a stable and  fast hierarchical LLM-powered multi-agent system. Our approach encompasses a  general framework with diverse communication patterns for fine-tuning and inference,  designed to achieve both high performance and efficiency.  The proposed tasks are as follows:  1) Designing an LLM-powered multi-agent architecture for solving complex tasks and  understanding communications between agents: We propose a hierarchical approach to  effectively organize role-based LLMs with diverse expertise. Intelligent agents will be  clustered and ensembled to produce sub-task outputs, and the final results will be  achieved through collaborations.  2) Designing efficient communication approaches for transmitting various patterns  among agents: This work focuses on three types of communication patterns—bits,  gradients, and logits—to fit different scenarios. Bit communication enables prompt  sharing among agents, using an LLM-based compressor to encode texts and accelerate  transmission. Gradient communication supports weight exchange during role-based  LLM fine-tuning via stochastic lattice and vector quantizers. While logit communication  uses logit ensembling techniques for effective model collaboration. These methods,  applicable beyond our framework, improve efficiency in diverse multi-agent  environments.  3) Applying and evaluating our proposed framework in complex tasks, like AI for math  and world simulation. This practical assessment will demonstrate the real-world impact  of the developed strategies and technologies.  By combining expertise in information theory, networking, communication, distributed  learning, and natural language processing, this project aims to achieve a comprehensive  understanding of communications in the LLM-powered multi-agent framework. It  encompasses algorithmic design, performance analysis, and proof-of-concept  experiments, providing insights to further refine the framework.  We anticipate that this project will not only advance our theoretical understanding of the communication efficiency issue in multi-agent systems but will also provide  application solutions for small businesses and industries grappling with these challenges.  The research outcomes are expected to inspire and benefit the broader research  community, contributing to the ongoing evolution of related fields.