HMAW: Hierarchical Multi-Agent Workflow for Prompt Optimization

Yuchi Liu, Jaskirat Singh, Wen Gao, Ali Payani, Liang Zheng

This repository contains the implementation of the Hierarchical Multi-Agent Workflow (HMAW) for prompt optimization in large language models (LLMs).

---

Prerequisites

Before running the scripts, ensure you have Python installed along with the necessary packages. You can install these packages using the provided requirements file.

Required Packages

• fire==0.5.0
• ipython==8.12.3
• numpy==1.26.4
• openai==1.30.4
• pandas==2.2.2
• setuptools==68.2.2
• tqdm==4.66.1

🔧 Installation

To install the required packages, execute the following command:

pip install -r requirements.txt

🔧 vLLM Model Installation

For installing and running the Mixtral model using vLLM, refer to the official vLLM documentation.

---

Directory Structure Preparation

To utilize the provided scripts effectively, please organize your data according to the following directory structure:

├── data
│   ├── Education
│   ├── FED
│   ├── GSM8K
│   ├── ATLAS
│   └── CodeNet
├── log
│   ├── compare_ATLAS_HMAW_base_mixtral_gpt3.5.json
│   ├── response_ATLAS_HMAW_mixtral.json
│   ├── response_ATLAS_base_mixtral.json
│   ├── compare_Education_HMAW_base_mixtral_gpt3.5.json
│   ├── response_Education_HMAW_mixtral.json
│   ├── response_Education_base_mixtral.json
│   └── ...
├── eval_ATLAS.py
├── ...
├── eval_compare.py
├── eval_correctness.py
└── ...

Log Files Description

• compare_ATLAS_HMAW_base_mixtral_gpt3.5.json: Contains comparison results of evaluator preferences between responses generated using HMAW prompts and basic user queries. The mixtral model is the LLM agent, and gpt3.5 is the evaluator.

• response_Education_HMAW_mixtral.json: Contains responses generated using the HMAW method with the mixtral model as the agent.

Note

• The log directory contains detailed comparison and response logs for various models and methods.
• The root directory includes essential evaluation scripts (eval_ATLAS.py, eval_compare.py, eval_correctness.py, etc.).

By maintaining this structure, you ensure that the scripts can locate and process the data and logs correctly, facilitating seamless evaluation and comparison of model responses.

---

How to Run

🔧 Generate Responses

To execute prompt optimization and response generation, use the following command:

python eval_{dataset}.py

Replace {dataset} with one of the following options to represent different datasets: ATLAS, education, fed, codenet, or GSM.

🔧 Test

Testing on ATLAS, education, fed, and codenet:

Run the following command:

python eval_compare.py

Please ensure to update the related file names in eval_compare.py to test under various settings.

Testing on GSM：

Run the following command:

python eval_correctness.py

Customizing Prompting Methods

For details on how to change the prompting methods used in response generation, please refer to the prompter.py file.

---

Citation

If you find our code helpful, please consider citing our paper:

@misc{liu2024hierarchical,
      title={Towards Hierarchical Multi-Agent Workflows for Zero-Shot Prompt Optimization}, 
      author={Yuchi Liu and Jaskirat Singh and Gaowen Liu and Ali Payani and Liang Zheng},
      year={2024},
      eprint={2405.20252},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

License

This project is open source and available under the MIT License.