News

• [2026.01.26] Our paper has been accepted by ICLR 2026 ✨.
• [2025.01.26] Our project introduction has been featured on DeepWiki.
• [2025.06.11] Our post on X (aka. Twitter) has received many likes.
• [2025.06.11] We were featured as HuggingFace Daily Paper #3.

Overview

Reinforced Rule-based Reasoning (RuleReasoner) is a simple yet effective method enabling small reasoning models (SRMs) to effectively learn rule-based reasoning. Unlike large models that need complex training, RuleReasoner uses a curated collection of tasks and a domain-aware dynamic sampling approach, adjusting training based on historical performance. This simple yet effective technique allows SRMs to outperform frontier Large Reasoning Models (LRMs) by +4.1% on in-distribution tasks and +10.4% on out-of-distribution tasks, while also being more computationally efficient.

• Domain-aware dynamic sampling with higher training sampling efficiency and domain performance balance.

  

• Comprehensive Data curation for data curricula on rule-centric application.

  

• Rule Reasoner (8B and 4B) depicts comparable performance versus a wide range of baselines.

  

• Rule Reasoner (8B and 4B) also achives strong OOD performance across three benchmarks (subsets of rule-based reasoning) including BBH, ProverQA, and BBEH.

  

Table of Contents

• News
• Overview
• Quick Start
  • Prerequisites
  • Installation
  • Training
  • Evaluation
• Project Structure
• Contributing
• License
• Citation

Quick Start

Prerequisites

Running RuleReasoner requires the dependencies listed in requirements.txt.

Installation

Build RuleReasoner from the source and install dependencies:

1. Clone the repository:

   git clone https://github.com/bigai-nlco/RuleReasoner.git
   

2. Navigate to the project directory:

   cd RuleReasoner
   

3. Install the dependencies:

   pip install -r requirements.txt
   pip install -e ./verl
   pip install -e .
   

Training

Run the training with:

./scripts/train/train_mix.sh

Evaluation

Run the evaluation with:

./scripts/eval/eval_model.sh \
    --model $MODEL_PATH \
    --datasets $DATASET_PATH \
    --output-dir $OUTPUT_DIR

Project Structure

└── RuleReasoner
    ├── LICENSE
    ├── README.md
    ├── requirements.txt
    ├── scripts
    │   ├── build_dataset.py
    │   ├── data
    │   ├── eval
    │   └── train
    ├── setup.py
    ├── src
    │   ├── __init__.py
    │   ├── data
    │   ├── globals.py
    │   ├── system_prompts.py
    │   └── utils.py
    └── verl
	└── ...

Contributing

• Discussions: Share your insights, provide feedback, or ask questions in the discussions section.
• Issues: Submit bugs found or log feature requests for the RuleReasoner project in the issues section.
• Pull Requests: Review open PRs, and submit your own PRs.

Contributing Guidelines

1. Fork the Repository: Start by forking the project repository to your own account.
2. Clone Locally: Clone the forked repository to your local machine.

   git clone https://github.com/<YOUR-USERNAME>/RuleReasoner.git
   

3. Create a New Branch: Always work on a new branch, giving it a descriptive name.

   git checkout -b new-feature-x
   

4. Make Your Changes: Develop and test your changes locally.
5. Commit Your Changes: Commit with a clear message describing your updates.

   git commit -m 'Implemented new feature x.'
   

6. Push to your fork: Push the changes to your forked repository.

   git push origin new-feature-x
   

7. Submit a Pull Request: Create a PR against the original project repository. Clearly describe the changes and their motivations.
8. Review: Once your PR is reviewed and approved, it will be merged into the main branch.

License

Rulereasoner is distributed under the terms of the MIT License.

Citation

@inproceedings{
    liu2026rulereasoner,
    title={RuleReasoner: Reinforced Rule-based Reasoning via Domain-aware Dynamic Sampling},
    author={Yang Liu and Jiaqi Li and Zilong Zheng},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026},
    url={https://openreview.net/forum?id=MQV4TJyqnb}
}