23 May 2024 | Xudong Guo, Kaixuan Huang, Jiale Liu, Wenhui Fan, Natalia Vélez, Qingyun Wu, Huazheng Wang, Thomas L. Griffiths, Mengdi Wang
This paper explores the integration of prompt-based organizational structures to enhance the cooperation of embodied Large Language Models (LLMs) in multi-agent systems. Inspired by human organizations, the authors introduce a framework that imposes structured communication patterns on LLM agents to mitigate issues such as over-reporting and redundant information. Through experiments with embodied LLM agents and human-agent collaboration, the study highlights the impact of designated leadership on team efficiency and the spontaneous cooperative behaviors of LLM agents. The authors also develop a *Criticize-Reflect* framework to improve organizational prompts, leading to novel and effective team structures that reduce communication costs and enhance efficiency. The main contributions include a novel multi-LLM-agent architecture, a *Criticize-Reflect* framework for prompt optimization, and extensive experimental results demonstrating the benefits of hierarchical organization. The findings suggest that designated leaders and efficient organizational structures significantly improve team performance, aligning with existing literature on human organizations.This paper explores the integration of prompt-based organizational structures to enhance the cooperation of embodied Large Language Models (LLMs) in multi-agent systems. Inspired by human organizations, the authors introduce a framework that imposes structured communication patterns on LLM agents to mitigate issues such as over-reporting and redundant information. Through experiments with embodied LLM agents and human-agent collaboration, the study highlights the impact of designated leadership on team efficiency and the spontaneous cooperative behaviors of LLM agents. The authors also develop a *Criticize-Reflect* framework to improve organizational prompts, leading to novel and effective team structures that reduce communication costs and enhance efficiency. The main contributions include a novel multi-LLM-agent architecture, a *Criticize-Reflect* framework for prompt optimization, and extensive experimental results demonstrating the benefits of hierarchical organization. The findings suggest that designated leaders and efficient organizational structures significantly improve team performance, aligning with existing literature on human organizations.