Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

TL;DR

This work tackles the dynamic design of communication topologies for LLM-driven multi-agent systems by introducing Guided Topology Diffusion (GTD), a conditional discrete graph-diffusion framework that generates task-specific topologies conditioned on context $C$. GTD integrates a lightweight surrogate reward model with a diffusion-based graph generator and employs zeroth-order guidance to steer per-step denoising toward high-reward topologies, balancing multi-objective criteria such as Utility, Token Cost, Sparsity, and Robustness. Empirically, GTD achieves state-of-the-art or highly competitive task performance across GSM8K, MATH, MultiArith, and SVAMP, while significantly reducing communication tokens and exhibiting strong robustness to agent failures. The combination of diffusion-based topology synthesis and gradient-free, proxy-guided optimization offers a scalable, data-efficient approach to Pareto-optimal topology design, with practical implications for energy-efficient and reliable multi-agent reasoning with LLMs.

Abstract

The efficiency of multi-agent systems driven by large language models (LLMs) largely hinges on their communication topology. However, designing an optimal topology is a non-trivial challenge, as it requires balancing competing objectives such as task performance, communication cost, and robustness. Existing frameworks often rely on static or hand-crafted topologies, which inherently fail to adapt to diverse task requirements, leading to either excessive token consumption for simple problems or performance bottlenecks for complex ones. To address this challenge, we introduce a novel generative framework called \textit{Guided Topology Diffusion (GTD)}. Inspired by conditional discrete graph diffusion models, GTD formulates topology synthesis as an iterative construction process. At each step, the generation is steered by a lightweight proxy model that predicts multi-objective rewards (e.g., accuracy, utility, cost), enabling real-time, gradient-free optimization towards task-adaptive topologies. This iterative, guided synthesis process distinguishes GTD from single-step generative frameworks, enabling it to better navigate complex design trade-offs. 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 in LLM agent collaboration.

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

TL;DR

This work tackles the dynamic design of communication topologies for LLM-driven multi-agent systems by introducing Guided Topology Diffusion (GTD), a conditional discrete graph-diffusion framework that generates task-specific topologies conditioned on context . GTD integrates a lightweight surrogate reward model with a diffusion-based graph generator and employs zeroth-order guidance to steer per-step denoising toward high-reward topologies, balancing multi-objective criteria such as Utility, Token Cost, Sparsity, and Robustness. Empirically, GTD achieves state-of-the-art or highly competitive task performance across GSM8K, MATH, MultiArith, and SVAMP, while significantly reducing communication tokens and exhibiting strong robustness to agent failures. The combination of diffusion-based topology synthesis and gradient-free, proxy-guided optimization offers a scalable, data-efficient approach to Pareto-optimal topology design, with practical implications for energy-efficient and reliable multi-agent reasoning with LLMs.

Abstract

The efficiency of multi-agent systems driven by large language models (LLMs) largely hinges on their communication topology. However, designing an optimal topology is a non-trivial challenge, as it requires balancing competing objectives such as task performance, communication cost, and robustness. Existing frameworks often rely on static or hand-crafted topologies, which inherently fail to adapt to diverse task requirements, leading to either excessive token consumption for simple problems or performance bottlenecks for complex ones. To address this challenge, we introduce a novel generative framework called \textit{Guided Topology Diffusion (GTD)}. Inspired by conditional discrete graph diffusion models, GTD formulates topology synthesis as an iterative construction process. At each step, the generation is steered by a lightweight proxy model that predicts multi-objective rewards (e.g., accuracy, utility, cost), enabling real-time, gradient-free optimization towards task-adaptive topologies. This iterative, guided synthesis process distinguishes GTD from single-step generative frameworks, enabling it to better navigate complex design trade-offs. 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 in LLM agent collaboration.

Paper Structure

This paper contains 37 sections, 4 theorems, 24 equations, 9 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Communication Topologies in Multi-Agent Systems
  4. Dynamic Topology Generation for LLM Agents
  5. Graph Diffusion Models for Synthesis
  6. Preliminaries
  7. Formalizing Topology Generation as a Conditional Generative Problem
  8. Optimization Objective.
  9. Denoising Diffusion Models for Graph Generation
  10. Forward Diffusion Process.
  11. Learned Reverse Process.
  12. The Challenge: Guiding Generation with a Black-Box Objective
  13. Methodology
  14. Surrogate Reward Model
  15. Architecture.
  16. ...and 22 more sections

Key Result

Lemma C.2

The posterior distribution $q(A_{t-1}|A_t, A_0)$ is a Gaussian distribution given by: where $\tilde{\boldsymbol{\mu}}_t(A_t, A_0) = \frac{\sqrt{\bar{\alpha}_{t-1}}\beta_t}{1-\bar{\alpha}_t}A_0 + \frac{\sqrt{\alpha_t}(1-\bar{\alpha}_{t-1})}{1-\bar{\alpha}_t}A_t$ and $\tilde{\beta}_t = \frac{1-\bar{\alpha}_{t-1}}{1-\bar{\alpha}_t}\beta_t$.

Figures (9)

  • Figure 1: Comparison of Multi-Agent System (MAS) communication topology design workflows. (1) Static Fixed Workflow, (2) Centralized Adaptive Workflow, (3) Diffusion Guided Topology Workflow (Ours). Our proposed method provides task- and context-adaptive topologies by using a conditional diffusion process guided by a proxy model to jointly optimize for utility, cost, robustness, and sparsity.
  • Figure 2: The Guided Topology Diffusion (GTD) framework workflow, divided into four main stages. 1) Material: The process begins with task-specific inputs, including the query, available agents, and tools. 2) Dataset Generation: A multi-agent framework simulates various baseline topologies to generate a foundational dataset linking topologies to performance outcomes (e.g., utility and cost). 3) Model Training: The generated dataset is used to train two core components: a lightweight proxy scorer ($P_{\phi}$) to predict topology performance and a conditional graph diffusion generator ($G_{\theta}$) to learn the structure of high-performing graphs. 4) Inference: For a new task, the framework uses the trained models to iteratively denoise a random graph, with the proxy scorer guiding each step to synthesize a final, task-optimized topology.
  • Figure 3: An illustration of different multi-agent communication topologies. The left panel shows examples of common static or heuristic graphs, such as Chain, Star, Complete, Layered, and Random graphs. The right panel shows examples of Adaptive Graphs, which represent the sparse, task-specific topologies that the GTD framework is designed to generate dynamically.
  • Figure 4: Accuracy versus token consumption for various multi-agent methods across the GSM8K, MultiArith, MMLU, and SVAMP benchmarks. The plots illustrate that topologies generated by GTD are highly cost-efficient, achieving strong performance while using significantly fewer tokens than baseline methods that rely on dense communication graphs.
  • Figure 5: Robustness of various multi-agent systems to simulated agent failure on the GSM8K benchmark. The chart compares task accuracy before and after an attack, demonstrating that topologies generated by GTD exhibit greater resilience and more graceful performance degradation compared to other methods.
  • ...and 4 more figures

Theorems & Definitions (9)

  • Definition C.1: Evidence Lower Bound (ELBO)
  • Lemma C.2: Forward Process Posterior
  • Theorem C.3: Optimality of the Denoising Objective
  • proof
  • Definition C.4: $\epsilon$-Accurate Surrogate Model
  • Theorem C.5: Performance Gap Bound
  • proof
  • Corollary C.6: Perfect Surrogate
  • Definition C.7: ZO-Guided Denoising Step