💬 NLP

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

This paper introduces Guided Topology Diffusion (GTD), a novel generative framework that leverages conditional discrete graph diffusion models and a lightweight proxy model to dynamically synthesize task-adaptive, efficient communication topologies for multi-LLM agent systems, thereby overcoming the limitations of static designs and significantly improving collaboration performance.

Original authors: 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

Published 2026-05-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: 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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a team of expert AI assistants (like a group of brilliant but very chatty consultants) trying to solve a difficult puzzle together. The big question is: How should they talk to each other?

In the past, researchers tried two main approaches:

The "One-Size-Fits-All" Rule: Everyone talks to everyone (like a chaotic town hall meeting). This works for hard problems but wastes a lot of time and money on simple ones.
The "Fixed Script": Everyone follows a strict line (like an assembly line). This is efficient but breaks down if the problem gets too complex or changes unexpectedly.

The paper introduces a new method called Guided Topology Diffusion (GTD). Here is how it works, using simple analogies:

1. The Problem: Designing the Conversation Map

Think of the team's communication structure as a road map.

If the map is too crowded (everyone connected to everyone), you get traffic jams (too many messages, high cost).
If the map is too sparse (only one person talks to the next), you get bottlenecks (the team can't solve hard problems).
Existing methods usually pick a map before they know the specific destination. They might use a "star" map for everything, which is inefficient.

2. The Solution: A "Sculptor" that Carves the Map

The authors propose a new way to create the perfect map for each specific task on the fly. They use a technique called Diffusion, which you can think of like sculpting from a block of clay.

The Starting Point (The Clay): Imagine starting with a messy, noisy block of clay where every possible connection between agents is randomly jumbled. It's a chaotic mess.
The Process (Carving): Instead of building the map piece by piece, the system starts with the mess and slowly "carves" away the bad connections, step-by-step, until a clean, efficient shape remains.
The Guide (The Sculptor's Eye): This is the magic part. Usually, a sculptor just guesses what the final statue should look like. In this system, there is a lightweight "Proxy" model acting as the sculptor's eye.
- At every single step of the carving, the Proxy looks at the current shape and asks: "Is this shape going to be accurate? Is it cheap to run? Is it robust?"
- If the shape is getting too expensive, the Proxy says, "Cut that connection."
- If the shape is too weak, the Proxy says, "Add a bridge here."

3. Why This is Different

Most other methods try to guess the perfect map in one giant leap (like throwing a dart at a board and hoping it hits the bullseye). If they miss, they have to start over.

GTD is different because it takes many small, guided steps.

Analogy: Imagine trying to find the best route through a foggy forest.
- Old Method: You guess the whole path at once. If you guess wrong, you get lost.
- GTD Method: You take one step, check a compass (the Proxy), adjust your direction, take another step, check again, and adjust. By the time you reach the end, you have found the perfect, most efficient path that balances speed, safety, and cost.

4. The Results

The authors tested this "sculpting" method on various tasks, like solving math problems and writing code.

Better Performance: The teams using these custom-made maps solved problems more accurately than teams using fixed maps.
Cheaper: Because the system "carves away" unnecessary connections, the teams sent far fewer messages (tokens), saving money and time.
Stronger: If one agent in the team fails or gives a wrong answer, the GTD-designed map has just enough backup connections to keep the team working, whereas rigid maps often collapse.

Summary

The paper presents a system that doesn't just pick a pre-made communication style. Instead, it dynamically sculpts a unique conversation map for every single task, guided by a smart assistant that constantly checks if the map is efficient, accurate, and strong. The result is a team of AI agents that works together perfectly, spending only as much energy as the job requires.

Technical Summary: Dynamic Generation of Multi-LLM Agent Communication Topologies with Graph Diffusion Models

1. Problem Statement

The efficiency of Multi-Agent Systems (MAS) driven by Large Language Models (LLMs) is fundamentally constrained by their communication topology. Current approaches predominantly rely on static, hand-crafted patterns (e.g., chains, stars, or fully connected graphs) or heuristic workflow templates. These fixed designs fail to adapt to the intrinsic complexity of specific tasks or the real-time progress of agent collaboration. Consequently, simple tasks often incur excessive token consumption due to unnecessary communication, while complex tasks suffer from performance bottlenecks due to insufficient collaboration structures.

Existing adaptive methods attempt to optimize topologies but face two critical limitations:

Generative Constraints: Many rely on single-step generation mechanisms (e.g., Variational Auto-Encoders), which struggle to capture the long-range dependencies inherent in optimal communication structures, often yielding topologies that are plausible but not Pareto-optimal.
Optimization Granularity: Optimization is often coarse-grained, applying reward signals only after a complete topology is generated. This post-hoc guidance makes it difficult to navigate the intricate, fine-grained trade-offs between competing objectives such as task utility (accuracy), communication cost (token consumption), and robustness (resilience to agent failure).

The core challenge is to develop a framework capable of precisely constructing task-adaptive topologies by integrating multi-objective guidance directly into the generative process, balancing utility, cost, and robustness without relying on differentiable reward functions.

2. Methodology: Guided Topology Diffusion (GTD)

The authors propose Guided Topology Diffusion (GTD), a novel framework that reframes topology synthesis as a conditional, iterative construction process using discrete graph diffusion models. GTD integrates a lightweight proxy model to steer the generation trajectory in real-time.

2.1 Core Components

Conditional Discrete Graph Diffusion: The framework models the communication topology as a directed graph $G=(V, E)$ represented by an adjacency matrix $A$ . It employs a variance-preserving forward diffusion process that gradually adds Gaussian noise to a clean graph $A_0$ over $T$ timesteps. The generative model learns the reverse process $p_\theta(A_{t-1}|A_t, C)$ , conditioned on a task-specific context vector $C$ (comprising the query and available agents), to denoise the graph back to a high-performance structure.
Surrogate Reward Model ( $P_\phi$ ): To address the high computational cost of evaluating topologies via full multi-agent simulations, GTD trains a lightweight Graph Neural Network (GNN) with Graph Attention (GAT) layers. This surrogate model predicts the performance vector (utility and cost) for a given graph-condition pair, approximating the expensive simulation outcomes.
Proxy-Guided Zeroth-Order Optimization: Since the true reward function is non-differentiable (a "black-box" objective) and the surrogate model operates on discrete graphs, standard gradient-based guidance is inapplicable. GTD employs a Zeroth-Order (ZO) optimization scheme during the reverse diffusion process. At each denoising step:
1. The generator predicts an unguided clean graph.
2. $K$ discrete candidate graphs are sampled from this prediction.
3. The surrogate model evaluates all candidates.
4. The candidate maximizing a composite reward (weighted utility minus cost) is selected.
5. This "best" candidate replaces the original prediction to compute the posterior distribution for the next timestep.

This iterative, guided synthesis allows the framework to navigate the design space toward Pareto-optimal trade-offs without requiring end-to-end differentiability.

3. Key Contributions

The paper outlines three primary contributions:

Problem Level: The introduction of GTD, a novel conditional discrete graph diffusion framework specifically designed for the dynamic generation of multi-agent communication topologies.
Algorithm Level: The design of a proxy-model-based zeroth-order optimization guidance algorithm. This effectively optimizes non-differentiable, high-cost external objectives (utility, cost, robustness) during the diffusion sampling process.
Framework Level: The construction of a complete end-to-end solution integrating semantic feature encoding, conditional graph diffusion, and multidimensional protocol-based dynamic guidance, offering a new paradigm for complex graph generation problems.

4. Experimental Results

The authors validated GTD across multiple benchmarks (GSM8K, MATH, MultiArith, HumanEval, MMLU, SVAMP) using GPT-4o-mini and Qwen-3-8B backbones.

Task-Solving Effectiveness: GTD achieved state-of-the-art performance across all tested benchmarks. For instance, on the challenging MATH dataset, GTD improved accuracy by over 2 percentage points compared to the strongest baseline (MaAS), reaching 54.07% accuracy. On GSM8K, it achieved 94.14% accuracy, significantly outperforming existing methods.
Communication Cost-Efficiency: GTD generates significantly sparser topologies. On GSM8K, it achieved >94% accuracy while consuming only ~4.8 million tokens, whereas comparable methods required 15% more tokens for lower accuracy, and some baselines used over five times the tokens. The framework successfully learns to preserve only critical communication links, avoiding the quadratic overhead of fully connected graphs.
Robustness: In simulations of agent failures on GSM8K, GTD-generated topologies demonstrated superior resilience. While other methods suffered significant accuracy drops (e.g., DyLAN dropped ~13 points), GTD's accuracy degraded by only 0.3 percentage points, indicating graceful degradation and effective redundancy.
Ablation Studies: Removing the proxy guidance mechanism caused a performance drop of nearly 6 percentage points on GSM8K, confirming the necessity of the guided synthesis. The framework also demonstrated high data efficiency, requiring only a minimal subset of 50 training samples to achieve significant gains.

5. Significance and Claims

The paper claims that GTD represents a significant advancement over static and single-step generative approaches by enabling fine-grained, multi-objective optimization of communication structures. By treating topology design as a guided diffusion process, the framework can dynamically adapt to task requirements, producing sparse, robust, and highly efficient collaboration networks.

The authors emphasize that while the current implementation requires a pre-computed seed dataset for training the surrogate model, the approach offers a principled way to navigate complex design trade-offs that static or rule-based systems cannot. The framework is presented as a new paradigm for solving complex graph generation problems in multi-agent systems, with the potential to scale to larger agent populations without hitting hardware bottlenecks. The work acknowledges limitations regarding the static nature of the generated topology (fixed before task execution) and the initial setup cost for data collection, positioning these as targets for future work in online active learning and dynamic evolution.