A Mixture-of-Agents System for Fact-Based Comprehensive Search

Abstract

Large Language Models (LLMs) exhibit remarkable capabilities; however, they possess inherent limitations due to static training, which leads to outdated information and hallucinations. Furthermore, most existing multi-agent frameworks depend on predefined and rigid agent roles that lack adaptability across a range of tasks. We introduce AMASS (Autonomous Multi-Agent System for Fact-Based Comprehensive Search), a dynamic and self-organizing framework that addresses these limitations through two primary innovations: autonomous agent generation and a Mixture of Agents (MoA) strategy. In contrast to traditional systems, AMASS spontaneously generates specialized agents tailored to the unique requirements of each task. Each agent autonomously selects the most appropriate LLM, based on the complexity of the task, thereby utilizing lightweight models for straightforward queries and more powerful models for intricate reasoning tasks. A centralized Critic Agent supervises the system to eliminate redundancy, ensure coherent collaboration, and continuously optimize the alignment of agents to tasks. Our cascaded search mechanism further enhances the accuracy of results by enabling an iterative and context-sensitive refinement process. Evaluated across the GPQA, Bamboogle, MuSiQue, and TriviaQA datasets, AMASS significantly outperforms state-of-the-art frameworks in accuracy, efficiency, and factual reliability, thus performing better for autonomous multi-agent reasoning systems.