File size: 964 Bytes
e0e9fc0
 
 
a16cc13
e0e9fc0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

<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>
  &nbsp;
  <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>
  &nbsp;
  <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>



![img](./exp1.png)