- [2025/05/06] We present the MAS-Zero [Project Page | Paper | Code]
- 🏠 [Project Page]
- 📜 [Paper]
- 💻 [Code]
We propose MAS-Zero, a meta-agent that serves several roles (design, evaluate, and verify) and involves two steps:
-
Meta-Iterations:
- MAS-Design: Task Decomposition and propose sub-MAS for each sub-task. We frame the MAS design as code generation.
- MAS-Feedback: Evaluate the generated MAS design on solvability and completeness. We evalaute these metrics using intermediate outputs by executing the MAS code.
- Self-Verification: selects the most suitable outcome from the set of all candidate solutions generated throughout the meta-iteration process.
In the whole process: no validation set needed; Meta-level self-supervision on MAS design; Inference-time only.
MAS-Zero sets a new frontier in the performance-cost trade-off across diverse domains and LLMs.
Our approach achieves strong performance across mathematical reasoning, graduate-level QA, and code benchmarks, using GPT-4o, LLaMA3.3-70B, and Qwen2.5-32B, without relying on any external supervision.
conda create -n mas_zero python=3.12 && conda activate mas_zero
pip install anthropic
pip install openai
pip install backoff
pip install together
cd ./
pip install -r requirements.txt
pip install datasets
pip install jinja2
⚠️ WARNING⚠️ : The implementation in this repository is very raw and intended for research purposes only. It is not secure for production environments. We plan to update our code to more secure implementations in the future. Your use of our code is at your own discretion and risk.
Search generates candidate answers. You can change AIME (aime24) to GPQA (gpqa_diamond) or SWE-Bench (swe_bench). For SWE-Bench, you need to follow the SWE-Bench instructions to install the Docker environment first. You can also modify meta_model and node_model to other LLMs. Please refer to the sampler/ folder (we support GPT, Claude, VLLM, and TogetherAI).
export OPENAI_API_KEY={YourKey}
export TOGETHER_API_KEY={YourKey}
python main_question.py --dataset workflow_search/aime24 --option plan --meta_model gpt-4o_chatgpt --node_model gpt-4o_chatgpt --verifier_model gpt-4o_chatgpt --blocks COT COT_SC Reflexion LLM_debate --use_oracle_verifier --defer_verifier --n_generation 5
After the search, we will need to select the best answer from the set of candidate answers. This is the verification step. Similarly, you can change AIME (aime24) to GPQA (gpqa_diamond) or SWE-Bench (swe_bench). You can also modify model to other LLMs. Please refer to the sampler/ folder (we support GPT, Claude, VLLM, and TogetherAI).
export OPENAI_API_KEY={YourKey}
export TOGETHER_API_KEY={YourKey}
python main_judge.py --dataset aime24 --judge_method self --baseline workflow_search --model gpt-4o_chatgpt --min_sample 0 --max_sample 30 --max_response_per_sample 9
If you find MAS-Zero helpful, please consider starring this repo and citing our work. We would be very grateful!
@misc{ke2025maszero,
title={MAS-Zero: Designing Multi-Agent Systems with Zero Supervision},
author={Zixuan Ke and Austin Xu and Yifei Ming and Xuan-Phi Nguyen and Caiming Xiong and Shafiq Joty},
year={2025},
eprint={2505.14996},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2505.14996},
}This project received help from many researchers at Salesforce AI Research. The code is adapted from the ADAS. During development, we also referred to simple-evals, MaAS, and AFlow.
Many thanks to the authors of these projects for their excellent contributions!
Feel free to contact Zixuan Ke via email: zixuan.ke@salesforce.com