The paper introduces the Hierarchical Auto-Organizing System (HAS), a framework designed to enhance multi-agent navigation in the open-world environment of Minecraft. HAS leverages large language models (LLMs) to enable agents to autonomously organize and collaborate, addressing the challenges of efficient inter-agent communication and task distribution. The system features a hierarchical structure with centralized planning and decentralized execution, an auto-organizing mechanism for dynamic group adjustment, and a multi-modal memory to store and retrieve experiences. The authors evaluate HAS on various tasks, including multi-modal goal search, continuous block search, and map exploration, demonstrating superior performance compared to baselines. The results highlight the effectiveness of HAS in improving navigation efficiency and adaptability in complex environments.The paper introduces the Hierarchical Auto-Organizing System (HAS), a framework designed to enhance multi-agent navigation in the open-world environment of Minecraft. HAS leverages large language models (LLMs) to enable agents to autonomously organize and collaborate, addressing the challenges of efficient inter-agent communication and task distribution. The system features a hierarchical structure with centralized planning and decentralized execution, an auto-organizing mechanism for dynamic group adjustment, and a multi-modal memory to store and retrieve experiences. The authors evaluate HAS on various tasks, including multi-modal goal search, continuous block search, and map exploration, demonstrating superior performance compared to baselines. The results highlight the effectiveness of HAS in improving navigation efficiency and adaptability in complex environments.