PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2510.07799v2cs.CLcs.AI

TL;DR

Static or hand-crafted communication topologies do not adapt well to diverse task requirements, making topology design a multi-objective challenge. GTD uses proxy-guided conditional graph diffusion to iteratively construct task-adaptive topologies. The paper reports that it produces sparse, robust topologies that outperform existing methods.

  • Problem

    Designing communication topologies for LLM multi-agent systems requires balancing task performance, communication cost, and robustness, while many systems still use hand-crafted or heuristic patterns.

    Source: §1 · p. 1

  • Method

    GTD iteratively denoises a conditional discrete graph with lightweight proxy-reward guidance and gradient-free optimization during sampling.

    Source: §1 · p. 2, §3.3 · p. 4

  • Results

    GTD produces sparse, robust topologies that outperform existing methods.

    Source: §7 · p. 9

  • Takeaways & Limitations

    The paper concludes that conditional graph diffusion can construct adaptive communication networks for multi-agent systems.

    Source: §7 · p. 9

    The current framework requires pre-computed seed data and fixes the topology before a task, without supporting time-varying evolution during conversation.

    Source: A Limitations · p. 12

Abstract

from arXiv · show

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.

1 Introduction

Fixed or heuristic communication patterns can mismatch task needs, while prior adaptive methods often offer coarse guidance and limited exploration of competing objectives. GTD instead guides iterative graph diffusion with a proxy model to construct task-adaptive topologies.

1 Introduction

  • Static or rule-based topologies can waste communication on simple tasks or bottleneck complex ones, while existing searches often underweight cost, robustness, and sparsity.The introduction contrasts short linear exchanges for simple Q&A with richer collaboration networks for software development.Source: §1 · p. 1
  • Prior adaptive generators can miss complex long-range dependencies and apply reward signals only after topology generation, limiting fine-grained multi-objective guidance.The stated challenge is to integrate that guidance at each generation step.Source: §1 · pp. 1–2
  • GTD casts topology synthesis as conditional discrete graph diffusion and injects proxy-based, gradient-free multi-objective guidance throughout iterative denoising.The guidance steers generation in real time toward task-specific utility, communication cost, and robustness objectives.Source: §1 · p. 2
  • The framework integrates semantic feature encoding, conditional graph diffusion, and multidimensional protocol-based dynamic guidance into an end-to-end topology-generation approach.Figure 1 presents the workflow as task- and context-adaptive, jointly optimizing utility, cost, robustness, and sparsity.Source: §1 · p. 2, Figure 1 · p. 2

2 Related Work

Prior work links communication structure to system behavior and explores topology optimization for LLM-agent collaboration. GTD is positioned as combining proxy-guided zeroth-order optimization with conditional graph diffusion to steer generation toward multiple objectives.

2 Related Work

  • Classical MAS research connects communication topology to global behavior, including consensus speed and robustness, while star-network studies quantify propagation and single-point-failure trade-offs.Recent structural optimization work examines Pareto fronts balancing accuracy and resilience.Source: §2 · p. 2
  • LLM-agent topology research targets token overhead, collaboration schedules, and task-aware structures learned with GNNs or autoregressive models.These approaches also address scalability and latency in decentralized settings.Source: §2 · pp. 2–3
  • GTD is distinguished by integrating proxy-guided zeroth-order optimization during conditional graph-diffusion sampling to steer generation toward multi-objective optima.The related-work discussion contrasts this approach with prior progress in conditional graph diffusion and broader generative paradigms.Source: §2 · p. 3

3 Preliminaries

The preliminaries formulate topology design as conditional graph generation and introduce diffusion as a way to construct graphs iteratively. They then motivate proxy-guided, gradient-free steering because external rewards are costly and non-differentiable.

3.1 Formalizing Topology Generation as a Conditional Generative Problem

  • Topology generation is framed as choosing an adjacency matrix for a task and agent set to maximize a composite reward, with utility, communication cost, and sparsity among its criteria.Evaluating the reward requires a costly multi-agent simulation for each candidate graph.Source: §3.1 · p. 3

3.2 Denoising Diffusion Models for Graph Generation

  • Graph diffusion learns to reverse gradual noising: a context-conditioned denoiser predicts the clean graph, and iterative application generates a new topology from noise.The forward process corrupts the graph over discrete timesteps; the learned reverse process reconstructs it, conditioned on task context.Source: §3.2 · p. 3

3.3 The Challenge: Guiding Generation with a Black-Box Objective

  • A standard conditional diffusion model cannot explicitly optimize the external reward during generation, and the true reward is too slow and non-differentiable for direct sampling guidance.The denoiser's continuous output must be converted into a discrete graph for evaluation, breaking end-to-end differentiability.Source: §3.3 · pp. 3–4
  • GTD uses a lightweight reward proxy and zeroth-order optimization to guide iterative diffusion sampling toward task-specific multi-objective rewards.The framework workflow trains the proxy and graph generator from simulated baseline topologies, then uses the proxy during inference to guide denoising.Source: §3.3 · p. 4, Figure 2 · p. 4

4 Methodology

GTD combines a surrogate reward model with a task-conditioned graph diffusion generator, then uses surrogate guidance during iterative topology synthesis. The method trains on simulated topology outcomes and guides generation toward high-performing task-specific graphs.

4 Methodology

  • GTD uses a surrogate reward model and a conditional diffusion generator, trained on a precomputed dataset and integrated for guided synthesis at inference.The surrogate approximates simulation outcomes, while the generator learns the distribution of high-performing graph structures.Source: §4 · p. 4

4.1 Surrogate Reward Model

  • The surrogate predicts task utility and communication cost from a graph-condition pair, and its candidate rankings usually identify the true best graph.It is trained against simulated performance vectors; held-out evaluation reports low MSE for utility and cost and strong positive correlation for predicted and ground-truth cost.Source: §4.1 · pp. 4–5

4.2 Conditional Graph Diffusion Generator

  • The diffusion generator learns a task-conditioned distribution of high-quality topologies and iteratively refines graphs, allowing proxy guidance at each construction step.A Graph Transformer models global dependencies between edges, and training uses only graphs exceeding a performance threshold.Source: §4.2 · pp. 5–6

4.3 Proxy-Guided Topology Synthesis

  • At each denoising iteration, the surrogate scores sampled discrete candidate graphs, and the best-ranked candidate guides the next state using a zero-order step.The task query embedding and current graph-state embeddings form the context-aware condition; guidance effectiveness depends on surrogate fidelity.Source: §4.3 · pp. 6–7

5 Experiments

Experiments evaluate task-solving effectiveness, communication cost-efficiency, and resilience to agent failures. GTD achieves strong benchmark accuracy while using fewer tokens and degrading less under simulated failure.

5 Experiments

  • Evaluations use GPT-4o-mini agents and task-specific teams, with the proxy and diffusion models trained using 50 samples and inference guidance evaluating five candidate graphs per timestep.Teams comprise four MathSolver agents, four CodeSolver agents, or three KnowledgeableAcademic agents; the benchmarks span mathematics, coding, and science tasks.Source: §5 · p. 7

5.1 Task-Solving Effectiveness

  • GTD achieves state-of-the-art accuracy on GSM8K (94.14), MATH (54.07), MultiArith (98.88), and SVAMP (91.30), outperforming baselines on most benchmarks.On MATH, it improves on the strongest baseline, MaAS, by over 2 absolute percentage points.Source: §5.1 · p. 7

5.3 Robustness Against Agent Failures

  • Under simulated GSM8K agent failure, GTD accuracy drops 0.3 percentage points (94.1% to 93.8%), versus nearly 13 points for DyLan and over 2 for Complete Graph.The comparison assesses task accuracy before and after an attack; the reported results describe GTD as more resilient than the other methods.Source: Figure 4 · p. 8, §5.3 · p. 8

5.2 Communication Cost-Efficiency

  • Across GSM8K, MultiArith, MMLU, and SVAMP, GTD pairs strong accuracy with lower token use than dense-graph baselines, and its proxy-training cost is rapidly amortized.On GSM8K, GTD exceeds 94% accuracy with 4.8e+06 tokens; the one-time training-data cost of ≈4.0×10^5 tokens is rapidly amortized by ≈4.4×10^6 tokens saved per full-test run.Source: §5 · p. 7, Figure 5 · p. 8, §5.2 · p. 8

6 Ablation Studies

Ablations assess proxy guidance, team size, training-sample count, and generalization to other models and harder benchmarks. Proxy-guided synthesis is a key contributor, while GTD also performs strongly with Qwen-3-8B and on LiveCodeBench.

6.1 Hyperparameter

  • Removing proxy guidance drops GSM8K accuracy from 94.14% to 88.42%; replacing the proxy with random guidance yields only a minor improvement.The ablation identifies targeted proxy optimization as the key driver of success.Source: §6.1 · p. 9
  • Performance scales effectively up to four agents, with diminishing returns thereafter, while the largest gains from additional training samples occur within the first 50.The authors describe four agents as a trade-off between task performance and computational efficiency, and report effective training without a massive dataset.Source: §6.1 · p. 9

6.2 Generalization to Open-Source Models and Harder Benchmarks

  • With Qwen-3-8B, GTD achieves the highest average accuracy of 72.8% across GSM8K, MATH, and MMLU; on LiveCodeBench, it scores 30.8% Pass@1, above MaAS (29.3%) and the Base Model (25.4%).The authors report gains across a different backbone and task difficulty.Source: §6.2 · p. 9, Table 3 · p. 9

7 Conclusion

GTD is presented as a way to generate sparse, robust communication topologies that outperform existing methods. The authors identify data, timing, and team-size limitations, alongside possible dual-use and dataset-bias risks.

7 Conclusion

  • GTD produces sparse, robust topologies that outperform existing methods.The authors also state that the current framework requires seed data and does not yet support time-varying topology evolution.Source: §7 · p. 9

A Limitations

  • Topology is fixed before a task begins, and performance gains diminish beyond approximately four agents on current reasoning benchmarks.The method also incurs setup cost to collect the baseline-topology seed data used to train its surrogate reward model.Source: A Limitations · p. 12

B Ethics and Societal Impact

  • The authors flag dual-use risks, including potential coordination of disinformation campaigns or automated attacks, and warn that biased training data could yield suboptimal or inequitable structures.They also note potential benefits for scientific discovery and energy consumption.Source: B Ethics and Societal Impact · p. 12

C Computational Experiments

The section reports linear GPU-memory scaling in an agent-count analysis, gives experimental setup details, and outlines a variational justification for graph-diffusion training.

C Computational Experiments

  • 2.8 GB for 5 agents and 4.9 GB for 1000 agents were reported in the GPU-memory scalability analysis, which the authors describe as linear scaling.The experiments used a server with four NVIDIA A6000 GPUs, each with 48 GB of VRAM.Source: C Computational Experiments · pp. 12–13

D Detailed Experimental Setup

  • Experiments used GPT-4o-mini agents, with domain-specific team sizes, and the surrogate model used 50 training samples.The surrogate is a two-layer GAT network with hidden dimension 32; the diffusion generator is a two-layer, two-head Graph Transformer trained over 50 timesteps.Source: D Detailed Experimental Setup · p. 13

F Data Statistics

G Theoretical Justification

  • The theoretical section frames graph diffusion within variational inference and analyzes convergence properties of proxy-guided synthesis.Source: G Theoretical Justification · p. 13

G.1 Variational Perspective of Graph Diffusion

  • The variational account connects ELBO optimization to minimizing the KL divergence between the forward posterior and learned reverse process, yielding an L2 objective that predicts clean A0 from noisy At.With matched Gaussian variances, the KL reduces to a squared distance between means; the section then derives the simplified loss across timesteps.Source: §G.1 · pp. 13–15

G.2 Analysis of Proxy-Guided Synthesis

The analysis explains how proxy-guided denoising steers topology generation and relates surrogate accuracy to the generated topology’s reward. It also states a perfect-surrogate result and summarizes the inference-time comparison’s scope.

Proxy accuracy

  • The accuracy of the surrogate reward model directly bounds the suboptimality of the topology found by an ideal proxy-guided optimizer.Source: §G.2 · p. 15

Perfect-surrogate result

  • With a perfect surrogate, a topology maximizing proxy reward also maximizes the true reward.Source: §G.2 · p. 15

Inference-time comparison

  • Table 6 compares average per-instance inference time on GSM8K and its caption reports that GTD achieves the best accuracy while maintaining competitive speed.Source: Table 6 · p. 16

Proxy-guided synthesis

  • The zeroth-order step guides denoising by selecting the best-reward candidate among sampled graphs, biasing the sampling trajectory toward higher proxy reward.Using multiple samples reduces variance in the stochastic ascent-direction estimate, and the greedy stepwise update supports non-differentiable objectives.Source: §G.2 · p. 16

H Inference Time Analysis

On GSM8K, the passage reports GTD’s per-instance inference time and attributes its speed relative to most multi-agent baselines to sparse generated topologies and fewer LLM API calls.

Inference time

  • 7.9 s/inst. is the reported GTD inference time on GSM8K with four agents and one A6000 GPU.The time includes topology generation using approximately 50 diffusion steps with zeroth-order guidance, plus agent execution.Source: H Inference Time Analysis · p. 16
  • Sparse generated topologies reduce the number of LLM API calls, making GTD faster than most multi-agent baselines despite topology-generation overhead.Source: H Inference Time Analysis · p. 16

I Scalability Analysis

GTD’s measured GPU memory requirement scales linearly as agent count increases, remaining within standard consumer hardware capacity even for large swarms.

I Scalability Analysis

  • GTD’s GPU memory requirement scales linearly with agent count and remains within standard consumer hardware capacity for large swarms.The paper notes that its benchmarks focus on small teams of 4–5 agents, while the framework is technically capable of optimizing larger organizations without hardware bottlenecks.Source: I Scalability Analysis · p. 16, Table 7 · p. 16

J Surrogate Model Fidelity Analysis

The supplementary analyses assess surrogate fidelity, robustness, alternative generation strategies, candidate-count trade-offs, compute setup, LLM use, prompt-conditioned topologies, and agent roles.

J Surrogate Model Fidelity Analysis

  • On GTD-generated OOD topologies validated against LLM execution, the proxy achieves 72.8% Top-1 ranking accuracy and 89.3% Top-2 of 5.The authors describe this as a slight drop from in-distribution performance but sufficient for effective guidance.Source: J Surrogate Model Fidelity Analysis · p. 16

K Extended Robustness Analysis

  • The extended GSM8K robustness analysis evaluates accuracy degradation across failure scenarios, and its caption reports consistently more graceful degradation for GTD.The supplied passages do not give the individual scenario values.Source: K Extended Robustness Analysis · p. 16, Table 9 · p. 17

L Comparison with Alternative Generation Methods

  • On GSM8K, the table caption reports that GTD outperforms alternative graph-generation strategies, using fewer proxy evaluations than MCMC with competitive inference speed.The alternatives include direct GNN prediction and MCMC edge-flip Metropolis-Hastings sampling.Source: L Comparison with Alternative… · pp. 16–17, Table 11 · p. 17

M Ablation on Number of ZO Candidates (K)

N Computational Resources

  • Experiments, surrogate and diffusion model training, and multiagent simulations used four NVIDIA A6000 GPUs, each with 48GB of VRAM.Source: N Computational Resources · p. 17

O The Use of Large Language Models (LLMs)

  • Evaluated agents use GPT-4o-mini; manuscript-preparation LLM use was limited to language polishing and figure generation.The authors state that the research and intellectual content were completed by the authors.Source: O The Use of Large Language Models… · p. 17

P Prompts

Q Agent Roles and Descriptions

  • The framework’s agent roles span specialized perspectives, including knowledge extraction, search, critique, mathematics, and domain expertise.Figure 8 visualizes the role set; the roles provide diverse perspectives combined to produce the final answer.Source: Q Agent Roles and Descriptions · p. 17, Figure 8 · p. 19

Related papers

Loading 2510.07799v2…