iLearn-Lab/NeurIPS25-SymMPO

---
license: apache-2.0
---

<div align="center">
<h2 align="center"> <b>SymMPO: Mitigating Hallucination Through Theory-Consistent Symmetric Multimodal Preference Optimization</b>
</h2>

<a target="_blank" href="https://github.com/Liuwq-bit">Wenqi&#160;Liu</a>¹,
<a target="_blank" href="https://scholar.google.com/citations?user=29gP4okAAAAJ">Xuemeng&#160;Song</a>²,
<a target="_blank" href="https://scholar.google.com/citations?user=tvPOeFQAAAAJ">Jiaxi&#160;Li</a>³,
<a target="_blank" href="https://scholar.google.com/citations?user=im-bS2YAAAAJ">Yinwei&#160;Wei</a>¹,
<a target="_blank" href="https://scholar.google.com/citations?user=VWunnXEAAAAJ">Zheng&#160;Na</a>⁴,
<a target="_blank" href="https://scholar.google.com/citations?user=aZZfn90AAAAJ">Jianhua&#160;Yin</a>¹,
<a target="_blank" href="https://scholar.google.com/citations?user=yywVMhUAAAAJ">Liqiang&#160;Nie</a>⁵
<br>
¹Shandong University&#160;&#160;&#160;
²Southern University of Science and Technology&#160;&#160;&#160;
³University of Georgia&#160;&#160;&#160;
<br>
⁴National University of Singapore&#160;&#160;&#160;
⁵Harbin Institute of Technology, Shenzhen&#160;&#160;&#160;
<br />

<a href='https://arxiv.org/abs/2506.11712'><img src='https://img.shields.io/badge/Paper-Arxiv-red'></a>
<a href='https://huggingface.co/iLearn-Lab/NeurIPS25-SymMPO-7B'><img src='https://img.shields.io/badge/Model-7B-yellow'></a>
<a href='https://huggingface.co/iLearn-Lab/NeurIPS25-SymMPO-13B'><img src='https://img.shields.io/badge/Model-13B-yellow'></a>
<a href='https://huggingface.co/datasets/iLearn-Lab/NeurIPS25-SymMPO'><img src='https://img.shields.io/badge/Dataset-HF-blue'></a>

</div>

---

## Introduction

We present **SymMPO**, a framework for mitigating hallucination in multimodal large language models (MLLMs). Our method introduces a theory-consistent symmetric multimodal preference optimization approach that addresses the hallucination problem from a principled perspective. This repository provides the official implementation, pretrained checkpoints, and evaluation scripts built on top of [LLaVA](https://github.com/haotian-liu/LLaVA).

## Citation

If you find our work helpful, please consider citing:

```bibtex
@inproceedings{
liu2025mitigating,
title={Mitigating Hallucination Through Theory-Consistent Symmetric Multimodal Preference Optimization},
author={Wenqi Liu and Xuemeng Song and Jiaxi Li and Yinwei Wei and Na Zheng and Jianhua Yin and Liqiang Nie},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
year={2025},
url={https://openreview.net/forum?id=tIW29IpCwG}
}
```