Fan Nie, Lan Feng, Haotian Ye, Weixin Liang, Pan Lu, Huaxiu Yao, Alexandre Alahi, James Zou

Paper

Our W4S framework operates as an iterative process of workflow generation, execution, and refinement:

1. Workflow Generation: The weak meta-agent design a new workflow to leverage the given strong model, represented as executable Python code.
2. Execution and Feedback.: The generated workflow is executed by a strong model on validation samples, producing performance feedback.
3. Refinement: The meta-agent uses feedback to iteratively improve the workflow.

Store your API keys in key.env:

OPENAI_API_KEY=""
ANTHROPIC_API_KEY=""

conda create -n w4s python=3.11
conda activate w4s
pip install .

conda create -n w4s python=3.11
conda activate w4s
pip install .[vllm]

If you find this work useful, please cite our paper:

@misc{nie2025weakforstrong,
      title={Weak-for-Strong: Training Weak Meta-Agent to Harness Strong Executors}, 
      author={Fan Nie and Lan Feng and Haotian Ye and Weixin Liang and Pan Lu and Huaxiu Yao and Alexandre Alahi and James Zou},
      year={2025},
      eprint={2504.04785},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2504.04785}, 
}

We thank ADAS and AFlow for their codebase and prompts.