ReCode

Updating Code API Knowledge with Reinforcement Learning

📄arXiv • 🤗HuggingFace • 📖Datasets

Table of Contents

• 🌟Overview
• 🔧Installation
• 📚Dataset Preparation
• 📉Training
• 🚩Citation
• 🌻Acknowledgement

🌟Overview

Large Language Models (LLMs) exhibit remarkable code generation capabilities but falter when adapting to frequent updates in external library APIs. This critical limitation, stemming from reliance on outdated API knowledge from their training data, even with access to current documentation, impedes reliable code generation in dynamic environments. To tackle this issue, we propose ReCode (rule-based Reinforcement learning for Code Update), a novel framework that mimics human programmer adaptation to API changes.

🔧Installation

We recommend creating a new conda environment to run our project:

# create a conda environment
conda create -n recode python=3.10
conda activate recode

# clone our project
git clone https://github.com/zjunlp/ReCode.git
cd ReCode

# install dependencies
pip install -r requirements.txt

📚Dataset Preparation

We have uploaded our collected data to Hugging Face, you can download it here. Each piece of data contains the following six fields:

• dependency specifies the library;
• new_version indicates the required version of the library;
• description provides an explanation of the code's functionality;
• update_info notes the details of the updates
• old_code contains the original code snippets
• new_code contains the updated code snippets

The following illustrates an example:

{
    "dependency": "Numpy", 
    "new_version": "==2.2.0", 
    "description": "The code demonstrates how to assign a custom docstring to a NumPy ufunc to enhance its documentation and improve code readability.", 
    "update_info": "_add_newdoc_ufunc is now deprecated. ufunc.__doc__ = newdoc should be used instead.", 
    "old_code": "import numpy as np\n\nmy_ufunc = np.frompyfunc(lambda x: x**2, 1, 1)\n\nnp._add_newdoc_ufunc(my_ufunc, \"This is a custom ufunc that squares the input.\")", 
    "new_code": "import numpy as np\n\nmy_ufunc = np.frompyfunc(lambda x: x**2, 1, 1)\n\nmy_ufunc.__doc__ = \"This is a custom ufunc that squares the input.\""
}

📉Training

GRPO Training

For GRPO training, we provide the script located at scripts/grpo.sh. Before using it, simply fill in the necessary content within the script.

bash scripts/grpo.sh

DAPO Training

Due to the requirements of the verl library for data format, it is necessary to preprocess the data before training.

python3 src/DAPO/data_process.py

After that, you can directly execute the script:

bash scripts/dapo.sh

🚩Citation

If you find our work helpful, please cite our paper:

@misc{wu2025recodeupdatingcodeapi,
      title={ReCode: Updating Code API Knowledge with Reinforcement Learning}, 
      author={Haoze Wu and Yunzhi Yao and Wenhao Yu and Huajun Chen and Ningyu Zhang},
      year={2025},
      eprint={2506.20495},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.20495}, 
}

🌻Acknowledgement

Our code is build on trl and verl. Thanks to their great works!