MACOF: Multi-Agent Collaboration Optimization Framework

The remarkable performance of large language models (LLMs) based multi-agent in addressing multi-stage complex problems has garnered significant attention. However, multi-agent still face challenges stemming from the insufficient reasoning capabilities of LLMs and the inherent hallucinations within them, which lead to suboptimal quality of the generated content. In this paper, we develop an optimization framework MACOF: Multi-Agent Collaboration Optimization Framework to improve the output quality of multi-agent. The framework modules include a collective memory module, an assessment-reflection module, and a guidance module. Each agent iteratively reflects on the results of the assessment and guidance feedback and learns about the whole group through collective memory, optimizing the output and reducing internal illusions with the help of the guidance module. The collective memory simultaneously forms a local knowledge base with scored metrics, laying the foundation for optimizing and enhancing intelligent applications on the local end. This paper takes requirement modeling application as a case study to verify the effectiveness of the proposed method. The results show the proposed method also reduces the overall internal illusion error about 6% in the case study. This research provides an effetive optimization framework for enhancing the reliability and effectiveness of multi-agent.