| <h1 align="center">🛠️ ReAligner</h1> | |
| <p align="center"> | |
| <a href="https://arxiv.org/pdf/2506.12704"><img src="https://img.shields.io/badge/arXiv-arXiv%20Preprint-B31B1B?style=flat&logo=arxiv&logoColor=white" alt="arXiv Paper"></a> | |
| | |
| <a href="https://github.com/zwhong714/ReAligner"><img src="https://img.shields.io/badge/Homepage-Project%20Page-brightgreen?style=flat&logo=github" alt="Homepage"></a> | |
| | |
| <a href="https://huggingface.co/wh-zhu"><img src="https://img.shields.io/badge/Huggingface-Models-yellow?style=flat&logo=huggingface" alt="Models"></a> | |
| </p> | |
| <div> | |
| A flexible realignment framework is proposed to quantitatively control alignment during training and inference, combining Training-time Realignment (TrRa) and Inference-time Realignment (InRa). | |
| - We realign DeepScaleR-1.5B model and reduce token usage without performance loss and even enhance reasoning capabilities. | |
| </div> | |
| </div> | |
| <div> | |
| <br> | |
|  |