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

Researchers introduced Multi-Order Communication (MOC) for LLM-based multi-agent systems to better preserve multi-hop evidence during message passing.

MOC restructures and consolidates agent messages under token limits, using Semantic-Topological Merging to keep important information.

In tests across six datasets and multiple model backbones, MOC improved task performance while reducing communication cost.

The approach could make multi-agent LLM systems more accurate and efficient by improving information flow without increasing token usage.