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

Overview

MOC is a multi-order communication mechanism for LLM-based multi-agent systems that works with various agent topologies, such as random graphs, predefined graphs, and task-adaptive graphs, to enable multi-hop information exchange and reduce redundant message passing.

Quick Start

Install packages

conda create -n moc python=3.10
conda activate moc
pip install -r requirements.txt

LLM Setup

MOC supports two LLM backends: API-based models and local Ollama models.

API Mode

Add API keys in template.env and rename it to .env:

BASE_URL = ""
API_KEY = "" 

Local Mode (Ollama)

Install Ollama and pull the models:

ollama pull gemma2:27b  
ollama pull gemma2:9b    # distillation model

Run MOC on MMLU with Random Graph

python experiments/run_experiment.py --domain mmlu --mode Random --edge_density 0.7 --agent_nums 7 --random_dag_seed 42 --neighbor_hops 2

Acknowledgement

This code refers to GDesigner.