Multi-Agent Collaboration Frameworks for Task-Oriented Large Language Models
Abstract
The rapid advancement of Large Language Models (LLMs) has catalyzed significant progress in artificial intelligence, yet single-agent systems face inherent limitations in complex, multi-faceted task execution. This paper presents a comprehensive investigation into multi-agent collaboration frameworks specifically designed for task-oriented LLMs. We propose a novel hybrid architecture that integrates hierarchical coordination with decentralized peer-to-peer communication, enabling efficient task decomposition, specialized agent allocation, and collaborative problem-solving. Through extensive experimentation on realworld datasets including the HotpotQA benchmark and the MS MARCO passage ranking corpus, our framework demonstrates superior performance with an accuracy improvement of 16.2% over single-agent baselines and a 3.8% enhancement compared to existing multi-agent systems. We formulate the agent coordination problem mathematically, present an adaptive task allocation algorithm, and provide empirical evidence of scalability up to 12 collaborative agents. Our findings indicate that properly structured multi-agent systems can significantly enhance the reasoning capabilities, task completion rates, and overall reliability of LLM-based applications in complex domains.