Talk Structurally, Act Hierarchically: A Collaborative Framework for LLM Multi-Agent Systems
Zhao Wang, Sota Moriyama, Wei-Yao Wang, Briti Gangopadhyay, Shingo Takamatsu
2025-02-18
Summary
This paper talks about TalkHier, a new framework designed to improve how multiple AI agents work together on complex tasks by introducing better communication and problem-solving methods. It focuses on making these agents collaborate more effectively and produce accurate results.
What's the problem?
When AI systems with multiple agents try to work together, they often face issues like poor communication, incorrect outputs, or biases. These problems make it hard for the agents to handle complicated tasks that require teamwork and precise reasoning.
What's the solution?
The researchers developed TalkHier, which uses a structured communication protocol to help the AI agents share information more effectively. It also introduces a hierarchical refinement system that allows the agents to fix mistakes and improve their outputs over time. This approach was tested on various tasks like answering open-ended questions and creating advertisements, showing that TalkHier outperformed existing methods in accuracy and collaboration.
Why it matters?
This matters because it helps make AI systems with multiple agents more reliable and efficient, especially for solving complex problems. By improving how these agents communicate and refine their work, TalkHier could lead to better AI applications in areas like research, customer service, and creative industries.
Abstract
Recent advancements in LLM-based multi-agent (LLM-MA) systems have shown promise, yet significant challenges remain in managing communication and refinement when agents collaborate on complex tasks. In this paper, we propose Talk Structurally, Act Hierarchically (TalkHier), a novel framework that introduces a structured communication protocol for context-rich exchanges and a hierarchical refinement system to address issues such as incorrect outputs, falsehoods, and biases. TalkHier surpasses various types of SoTA, including inference scaling model (OpenAI-o1), open-source multi-agent models (e.g., AgentVerse), and majority voting strategies on current LLM and single-agent baselines (e.g., ReAct, GPT4o), across diverse tasks, including open-domain question answering, domain-specific selective questioning, and practical advertisement text generation. These results highlight its potential to set a new standard for LLM-MA systems, paving the way for more effective, adaptable, and collaborative multi-agent frameworks. The code is available https://github.com/sony/talkhier.