This paper has been accepted by ASE 2024 and awarded the 🏆 ACM SIGSOFT Distinguished Paper Award.

📖Overview • 🧪Environment • 📂Dataset Setup • 🚀Quick Start • 📝Citation

In this paper, we draw on pair programming practices to propose PairCoder, a novel LLM-based framework for code generation. PairCoder incorporates two collaborative LLM agents, namely a Navigator agent for high-level planning and a Driver agent for specific implementation.

The Navigator is responsible for proposing promising solution plans, selecting the current optimal plan, and directing the next iteration round based on execution feedback. The Driver follows the guidance of Navigator to undertake initial code generation, code testing, and refinement. This interleaved and iterative workflow involves multi-plan exploration and feedback-based refinement, which mimics the collaboration of pair programmers.

PairCoder is developed on Ubuntu 16.04 LTS. Please follow these steps to set up the Python environment:

conda create -n PairCoder python=3.10
conda activate PairCoder
pip install -r requirements.txt

Please set your API KEY in settings/configuration.toml. This file also contains numerous other configurable options that allow you to fine-tune and precisely control the behavior of PairCoder.

Download the dataset file from releases.

Unzip the file and place the dataset directory in the project root:

PairCoder/
├─assets
├─dataset
├─src

The dataset is required for running the experiments. Ensure that the directory structure matches the project requirements.

Use the following command to perform code generation:

python src/solve_dataset.py \
    --dataset_name mbpp \
    --split_name test \
    --dir_path results

For the given split_name of the dataset_name, the logs and the final solutions.json are stored in dir_path. You can set id_list for ids to solve.

If you find the code helpful, please cite our paper:

@inproceedings{zhang2024paircoder,
  title     = {A Pair Programming Framework for Code Generation via
                Multi-Plan Exploration and Feedback-Driven Refinement},
  author    = {Zhang, Huan and Cheng, Wei and Wu, Yuhan and Hu, Wei},
  booktitle = {The 39th IEEE/ACM International Conference on Automated Software Engineering (ASE 2024)},
  year      = {2024}
}