We propose AgentInit, a novel initialization framework for large language model (LLM)-based multi-agent systems. AgentInit strategically orchestrates diverse agent roles and expertise distributions to optimize collaboration effectiveness and efficiency.
- [20/08/2025] 🎉🎉Our paper is accepted by [EMNLP 2025 Findings]!🎉🎉
- [24/09/2024] Our paper is published on arXiv: arXiv:2509.19236!
Install anaconda environment
conda create -n agentinit python=3.10
conda activate agentinit
pip install -r requirements.txtSet url and API keys in AgentInit/llm/gpt_chat
MINE_BASE_URL = ""
MINE_API_KEYS = ""Set text encoder in AgentInit/agentinit/embedder.py
model_path: str = ''Prepare data from Huggingface. And put them in datasets/.
Run AgentInit on MMLU, the same as other datasets:
python -u experiments/run_mmlu.py --batch_size 10 --llm_name MODEL_NAME --agent_names AutoAgentRun on different framework:
python -u experiments_Autogen/run_mmlu.py --batch_size 10 ---llm_name MODEL_NAME You can also try our pre-generated examples to quickly verify the evaluation results without running the full pipeline:
python -u example/run_mmlu.py --batch_size 40 --llm_name YOUR QWEN MODEL --agent_names AutoAgentFor additional benchmarks and task-specific evaluation settings, please refer to the corresponding directories:
cd scienceworldcd writingIf you find this repo useful, please cite our paper as:
@inproceedings{
tian2025agentinit,
title={AgentInit: Initializing {LLM}-based Multi-Agent Systems via Diversity and Expertise Orchestration for Effective and Efficient Collaboration},
author={Chunhao Tian and Yutong Wang and Xuebo Liu and Zhexuan Wang and Liang Ding and Miao Zhang and Min Zhang},
booktitle={The 2025 Conference on Empirical Methods in Natural Language Processing},
year={2025},
url={https://openreview.net/forum?id=j9y7YaxoG0}
}
Code framework based on GPTSwarm, AgentPrune, AgentDropout and AutoAgents