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

Researchers propose Multi-Order Communication (MOC), a new framework for improving how large language model-based multi-agent systems exchange information. The scheme addresses limitations in current message-passing approaches by capturing multi-hop dependencies and consolidating messages efficiently, demonstrating consistent performance improvements across multiple datasets while reducing communication costs.

The advancement of multi-agent LLM systems has focused heavily on optimizing how agents coordinate their actions through various network topologies, yet a critical gap remains in how these agents actually communicate with one another. Current methods rely on concatenating first-order neighbor responses, creating a bottleneck where information becomes diluted as it travels through multiple agent hops and crucial insights get lost in translation. MOC addresses this fundamental communication problem by reconstructing how inter-agent messages flow, enabling agents to access information from multiple hops away while maintaining semantic clarity within computational constraints.

This research builds on the broader trend of improving LLM agent collaboration, which has become increasingly important as practical applications demand agents to solve complex problems requiring coordinated reasoning. The Multi-Order Communication scheme introduces a structured evidence stream that captures dependencies across multiple communication layers, paired with a Semantic-Topological Merging algorithm designed to preserve information quality while respecting token budget constraints—a practical consideration for production systems.

The implications extend beyond academic performance metrics. By reducing communication overhead while improving task performance, MOC addresses real scalability challenges that developers face when deploying multi-agent systems. The consistent improvements across six diverse datasets and LLMs of varying sizes suggests the approach is robust and generalizable. This is particularly valuable for enterprise applications where communication costs directly impact computational expenses and latency requirements. As LLM-based autonomous systems move toward real-world deployment, efficient communication architectures become critical infrastructure components that determine both system reliability and cost-effectiveness.

• →MOC reconstructs inter-agent communication to capture multi-hop dependencies, addressing limitations in direct message concatenation methods.
• →A Semantic-Topological Merging algorithm optimizes message quality while staying within token constraints.
• →Framework demonstrates consistent performance improvements and reduced communication costs across six datasets and multiple LLM scales.
• →Efficient agent communication architecture addresses critical scalability challenges for production multi-agent system deployment.
• →Research provides open-source implementation, enabling broader adoption and validation by the research community.