G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks

TL;DR

G-Designer introduces MACP, a protocol-guided approach to automatically design task-aware, efficient, and robust communication topologies for LLM-based multi-agent systems. By modeling agents as graph nodes and leveraging a variational graph auto-encoder, it constructs a task-specific, sparse G_com that balances performance with token efficiency and resilience to adversarial prompts. Extensive experiments across six benchmarks demonstrate strong, adaptable performance with favorable scalability and resource characteristics, outperforming many static or less adaptive topologies. The work highlights how task-aware topology design can substantially reduce communication overhead while maintaining high-quality solutions, advancing practical deployment of collaborative AI systems.

Abstract

Recent advancements in large language model (LLM)-based agents have demonstrated that collective intelligence can significantly surpass the capabilities of individual agents, primarily due to well-crafted inter-agent communication topologies. Despite the diverse and high-performing designs available, practitioners often face confusion when selecting the most effective pipeline for their specific task: \textit{Which topology is the best choice for my task, avoiding unnecessary communication token overhead while ensuring high-quality solution?} In response to this dilemma, we introduce G-Designer, an adaptive, efficient, and robust solution for multi-agent deployment, which dynamically designs task-aware, customized communication topologies. Specifically, G-Designer models the multi-agent system as a multi-agent network, leveraging a variational graph auto-encoder to encode both the nodes (agents) and a task-specific virtual node, and decodes a task-adaptive and high-performing communication topology. Extensive experiments on six benchmarks showcase that G-Designer is: \textbf{(1) high-performing}, achieving superior results on MMLU with accuracy at $84.50\%$ and on HumanEval with pass@1 at $89.90\%$; \textbf{(2) task-adaptive}, architecting communication protocols tailored to task difficulty, reducing token consumption by up to $95.33\%$ on HumanEval; and \textbf{(3) adversarially robust}, defending against agent adversarial attacks with merely $0.3\%$ accuracy drop.

Figures (6)

• Figure 1: Existing practices for LLM-based multi-agent communication topology design.
• Figure 2: The token consumption and accuracy of different multi-agent protocols on two subsets of MMLU dataset, "Highschool Biology" and "College Math", tested with four gpt-4-based agents.
• Figure 3: The designing workflow of our proposed G-Designer.
• Figure 4: Visualization of the performance metrics and prompt token consumption of different multi-agent communication topologies across MMLU, HumanEval, GSM8K, and SVAMP. The diameter of each point is proportional to its $y$-axis value.
• Figure 5: We compare the accuracy (%) of various multi-agent frameworks before and after prompt attacks on MMLU.

Theorems & Definitions (1)

• Definition 3.1: Multi-agent Communication Protocol