Sparse Communication Topologies for Multi-Agent Systems

This direction is split into other directions: sparse communication topology. Its last version stays readable and citable here.

Version 1 of 1187 papers256 conceptsfrozen 2026-10-04

In multi-agent systems (MAS), the default assumption is often that more communication is better. If agents can share every observation and every thought, they should, in theory, coordinate perfectly. However, recent research has revealed a counter-intuitive reality: "everyone-talks-to-everyone" architectures are not just economically expensive—they are often mathematically inferior.

Excessive communication creates a "free-for-all" bottleneck where the signal is drowned out by noise. As Jusheng Zhang and colleagues note in Agent-GSPO, unconstrained chatter degrades the signal-to-noise ratio, leading to worse performance than if agents spoke less. The emerging research direction focuses on Sparse Communication Topologies: the art of identifying and maintaining only the most critical interaction links between agents.

The Problem of Dense Communication

When every agent broadcasts to every other agent, two things happen. First, the cost (in tokens or bandwidth) scales quadratically with the number of agents. Second, agents suffer from information overload. Ziluo Ding and Zongqing Lu demonstrate in Learning Individually Inferred Communication for Multi-Agent Cooperation that sparse architectures prevent agents from processing irrelevant status updates, which is key to scaling systems to complex, high-density environments.

By filtering out irrelevant interactions, the learning signal becomes cleaner. Yong Liu shows in Multi-Agent Game Abstraction via Graph Attention Neural Network that this explicit noise removal leads to more stable training and lower variance compared to "soft" attention mechanisms, which often remain "leaky" and assign importance to distracting data.

Architecting Sparsity: Pruning and Generation

How do we find the "minimum viable" communication structure? Researchers have proposed two primary paths: Pruning existing dense networks and Generating sparse networks from scratch.

1. Strategic Pruning

Pruning treats the communication network as a graph and removes edges that contribute little to the final outcome. Guibin Zhang and the team behind AgentPrune found that reducing communication density by 10% to 30% can actually improve accuracy. Their framework slashes token usage by up to 72.8% by identifying redundant spatial-temporal links.

In multi-modal settings, this pruning can be hierarchical. Weiwei Shao introduces M3Prune, which first sparsifies connections within specific data modalities (like text or vision) before refining cross-modal links, ensuring that internal agent clusters are efficient before they attempt to coordinate across the whole network.

2. Generative Topology Design

Instead of starting with a mess and cleaning it up, some researchers build the "blueprint" for communication dynamically. Shiyuan Li proposes ARG-DESIGNER, which uses autoregressive graph generation to select agent roles and links based on the specific task query. Similarly, Eric Hanchen Jiang uses graph diffusion models in Guided Topology Diffusion to iteratively denoise a random graph into a task-specific communication network.

The Pareto Frontier of Coordination

The goal of these sparse topologies is to find the optimal balance between task accuracy and communication cost. Guibin Zhang demonstrates with G-Designer that sparse architectures can achieve over 90% reductions in token usage while maintaining or even exceeding the performance of dense networks.

This efficiency is often achieved through information-theoretic constraints. Tonghan Wang and Chongjie Zhang use the NDQ framework to apply regularizers that force agents to act independently most of the time, communicating only when coordination is critical. This "nearly decomposable" approach reduces overhead by 80% in complex environments like StarCraft II.

Structural Intuition: Small Worlds and Hierarchies

Why does sparsity work so well? One explanation comes from social and biological network theory. Boxuan Wang argues in Rethinking Multi-Agent Intelligence Through the Lens of Small-World Networks that a small number of strategic long-range connections can replace dense communication. This "Small-World" topology stabilizes consensus and reduces costs by mimicking the efficient wiring of the human brain.

Furthermore, Li Shen shows in CommFormer that treating communication as a learnable, sparse graph problem allows agents to filter out irrelevant noise while maintaining the coordination necessary for complex tasks. By explicitly defining the communication structure via a learnable adjacency matrix, systems can discover directed, task-aligned paths that avoid the "curse of dimensionality" inherent in all-to-all broadcasting.

Ultimately, the shift toward sparse topologies represents a move from "brute force" communication to "intelligent" coordination. As Jiaxi Yang and Srijan Kumar suggest in their position paper, the future of multi-agent systems lies in Topological Structure Learning—where the "who-talks-to-whom" is as carefully engineered as the agents themselves.

Papers in this version

Showing 60 of 187 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_sparse_communication_topologies_for_mult_fe747d5d,
  title = {Sparse Communication Topologies for Multi-Agent Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sparse-communication-topologies-for-multi-agent-systems/rdr_fe747d5d93fd4a128df35a53eb8dafe3},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_sparse_communication_topologies_for_mult_fe747d5d_v1,
  title = {Sparse Communication Topologies for Multi-Agent Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sparse-communication-topologies-for-multi-agent-systems/rdr_fe747d5d93fd4a128df35a53eb8dafe3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}