GTD: Dynamic Generation of Multi LLM Agents Communication Topologies with Graph Diffusion Models

Eric Hanchen Jiang, Guancheng Wan, Sophia Yin, Mengting Li, Yuchen Wu, Xiao Liang, Xinfeng Li, Yizhou Sun, Wei Wang, Kai-Wei Chang, Ying Nian Wu

Keywords: Graph, Agent, Multi-agent Systems, Large Language Model

Abstract: The efficiency of Multi-Agent Systems (MAS) largely hinges on their communication topology. Existing methods often rely on static or rule-based topologies, which struggle to meet diverse task requirements. To address this challenge, we introduce a novel generative framework called "Guided Topology Diffusion (GTD)." GTD is the first to apply a conditional discrete graph diffusion model for the dynamic generation of MAS communication topologies. Our core idea models the topology generation process as an iterative edge construction process, starting from an empty graph and guided by both task and agent team context. We designed a context-aware Graph Transformer as the denoising network and innovatively proposed a two-stage guidance mechanism: first, a lightweight Proxy Model quickly predicts non-differentiable, task-performance-related multidimensional communication protocols (such as utility, cost, robustness) as reward signals; then, during the sampling phase, a Zeroth-Order Optimization algorithm adjusts the generation trajectory in real-time based on this proxy reward. We validated GTD across multiple benchmarks, and experiments show that this framework can generate highly task-adaptive, sparse, and efficient communication topologies, significantly outperforming existing methods.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 10255