Title: MOC: Multi-Order Communication in LLM-based Multi-Agent Systems

Abstract:

While Large Language Model (LLM) driven Multi-Agent Systems have seen significant advancements, the majority of existing studies prioritize the optimization of coordination topologies. Consequently, the equally vital challenge of effectively transmitting and refining messages between agents remains largely unexplored. Standard communication protocols generally depend on the straightforward concatenation of responses from immediate neighbors. This approach creates a limited receptive field for evidence and causes essential insights to become diluted as they traverse multiple hops.

To overcome these constraints, we introduce the Multi-Order Communication (MOC) framework. This scheme reimagines inter-agent communication to capture multi-hop dependencies and integrates a structural message consolidation strategy to maintain operational efficiency. Our method formalizes the communication process to generate a structured, multi-order evidence stream. Furthermore, we developed a Semantic-Topological Merging algorithm designed to preserve semantic fidelity while adhering to token limits.

Extensive evaluations conducted across six varied datasets and LLM backbones with different parameter sizes show that MOC consistently enhances task performance while lowering communication expenses. The source code is accessible at https://github.com/yao-guan/MOC.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC