Authors: Taewon Yun, Jihwan Oh, Hyangsuk Min, Yuho Lee, Jihwan Bang, Jason Cai, Hwanjun Song

This is the official github repository for "ReFeed: Multi-dimensional Summarization Refinement with Reflective Reasoning on Feedback"

• 📌 Table of Contents
  • Overview
  • Model
  • Datasets
  • Quick Start
    • Training Setup
    • Launch Training
    • Inference
  • Results
  • Citation
  • Acknowledgement

ReFeed rethinks summary refinement by showing that reflective reasoning over multi-dimensional feedback is key to achieving balanced, robust improvements.

Our contributions:

• 🧠 Introduces reflective reasoning for multi-dimensional refinement, enabling models to resolve trade-offs, mitigate order bias, and filter noisy feedback simultaneously
• 🏗️ Releases SumFeed-CoT, a large-scale Long-CoT dataset that distills high-quality reflective reasoning from large reasoning models into lightweight models
• 🚀 Demonstrates strong empirical gains and robustness, achieving consistent improvements across faithfulness, completeness, and conciseness while remaining resilient to feedback quality and ordering

Our ReFeed model is available on Hugging Face 🤗:

We release our datasets through Hugging Face 🤗:

1. Environment Setup

   • We recommend following the environment setup in the open-r1 documentation on huggingface
   • Ensure all dependencies are properly installed and configured.

2. Data Preparation

   • Obtain the SumFeed-CoT dataset from 🤗 Hugging Face.

3. Configuration

   • Use our provided configuration file.

For training, use the following command:

sh ./script/sft.sh

For inference, use our provided prompt.

• 📈 ReFeed achieves the best overall performance across faithfulness, completeness, and conciseness, outperforming all previous refinement methods that optimize a single dimension.
• ⚖️ Reflective reasoning enables balanced improvements, effectively mitigating trade-offs that arise when optimizing multiple dimensions simultaneously.
• 🔀 Strong robustness to feedback order and noise, showing minimal performance variance under shuffled or low-quality feedback settings.
• 🧩 Efficient distillation: a lightweight 8B model matches teacher-level refinement quality while significantly reducing inference cost.

Please consider citation if our paper is useful in your research.

@inproceedings{yun2025refeed,
    title={ReFeed: Multi-dimensional Summarization Refinement with Reflective Reasoning on Feedback},
    author={Taewon Yun and Jihwan Oh and Hyangsuk Min and Yuho Lee and Jihwan Bang and Jason Cai and Hwanjun Song},
    booktitle={Second Conference on Language Modeling},
    year={2025},
    url={https://openreview.net/forum?id=6BGDGKZN7q}
}

This research was supported by KISTI, and by the NRF. For GPU infrastructure, our work was supported by the IITP grant funded by MSIT. This work was supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2025-02653113, High-Performance Research AI Computing Infrastructure Support at the 2 PFLOPS Scale)