Large Language Model-based Human-Agent Collaboration for Complex Task Solving

This repository is based on our paper: Large Language Model-based Human-Agent Collaboration for Complex Task Solving. It contains the human-agent collaboration dataset we generated, as well as demo code for our fine-tuned human-agent collaboration policy model.

Overview

• The Code for different datasets is in hotpotqa/, strategyqa/, and intercode/.
  • start training by scripts/run.sh
  • local test environment is in test_data/
• Human-Agent Collaboration Dataset in dataset/

Usage

Getting Start

You can use following scripts to install related python package through pip:

git clone https://github.com/XueyangFeng/ReHAC.git
cd ReHAC
pip install -r requirements.txt

Constructing Training Data

Here, we give an example where we set $\lambda=0.08$.

python data_preprocess.py ./dataset/gpt4/hotpotqa.jsonl 0.08 ./hotpotqa/data/advantage_sample_count_0.08.jsonl

The processed training data is then saved in ./hotpotqa/data/advantage_sample_count_0.08.jsonl and you should set TRAIN_DATA_DIR in run.sh to this path.

You can also find the training data we have processed under hotpotqa/data/advantage and strategyqa/data and intercode/data/sql folders.

Training and Test Process

cd hotpotqa/scripts
sh run.sh

Results

We random sample 100 questions for test for each dataset. The evaluation result of HotpotQA dataset is under the following figure:

(a) Human-agent collaboration evaluation. (b) GPT-4-agent collaboration evaluation. The bars below the 0-axis represent the human intervention cost $\lambda C$, the entire columns, composed of the bars above and below the 0-axis, represent the task reward $T$, and the bars above the 0-axis represent the reward $R$ ($R=T - \lambda C$). Numbers within the bars means the human intervention rate. $ReHAC_{GPT-4}$ and $ReHAC_{Human}$ represent the policy model trained on GPT-4-agent and human-agent collaboration datasets, respectively. ReHAC outperforms other baselines in human-agent collaboration scenarios.

We provide original evaluation outputs of ReHAC under hotpotqa/results, strategyqa/results, and intercode/results.

Citation

If you find that ReHAC is helpful for your work, please cite the follow paper.

@article{feng2024large,
  title={Large Language Model-based Human-Agent Collaboration for Complex Task Solving},
  author={Feng, Xueyang and Chen, Zhi-Yuan and Qin, Yujia and Lin, Yankai and Chen, Xu and Liu, Zhiyuan and Wen, Ji-Rong},
  journal={arXiv preprint arXiv:2402.12914},
  year={2024}
}