---
license: apache-2.0
base_model: Qwen/Qwen2.5-VL-7B-Instruct
tags:
- multimodal
- personalized-mllm
- ai-agent
- long-term-memory
- cvpr2026
- personality-evolving
- benchmark
datasets:
- ClareNie/Persona-MME
- ClareNie/PersonaVLM-Dataset
---

# PersonaVLM: Long-Term Personalized Multimodal LLMs (CVPR 2026)

<p align="center">
    <a href="https://PersonaVLM.github.io"><img src="https://img.shields.io/badge/Project-Page-blue.svg"></a>
    <a href="https://github.com/MiG-NJU/PersonaVLM"><img src="https://img.shields.io/badge/GitHub-Code-black"></a>
    <a href="http://arxiv.org/abs/2604.13074"><img src="https://img.shields.io/badge/arXiv-Paper-b31b1b.svg"></a>
    <img src="https://img.shields.io/badge/Accepted-CVPR%202026-red.svg">
</p>

> 🎉 **News:** Our paper "PersonaVLM: Long-Term Personalized Multimodal LLMs" is accepted to **CVPR 2026**! 

## 🌟 Introduction
**PersonaVLM** is an innovative personalized multimodal agent framework designed for long-term personalization. It transforms a general-purpose MLLM into a personalized assistant by integrating three key capabilities:
1. **Remembering**: Proactively extracts and summarizes multimodal memories into a personalized database.
2. **Reasoning**: Conducts multi-turn reasoning by retrieving relevant memories from a multi-type memory architecture (core, semantic, episodic, and procedural).
3. **Response Alignment**: Infers the user's evolving personality using a **Momentum-based Personality Evolving Mechanism (PEM)** to ensure aligned outputs.

## 📊 Persona-MME Benchmark
We establish **Persona-MME**, a comprehensive benchmark comprising over 2,000 curated interaction cases across 14 fine-grained tasks to assess long-term MLLM personalization.

## 🔗 Official Resources
This project consists of several components. You can access the model weights, training data, benchmark, and code via the links below:

| Resource | Link |
| :--- | :--- |
| 🌐 **Project Page** | [https://PersonaVLM.github.io](https://PersonaVLM.github.io) |
| 💻 **Official Code** | [GitHub: PersonaVLM](https://github.com/MiG-NJU/PersonaVLM) |
| 🤗 **Model Weights** | [HF: PersonaVLM (Qwen2.5-VL-7B)](https://huggingface.co/ClareNie/PersonaVLM) |
| 📊 **Benchmark** | [HF: Persona-MME (2,000+ cases)](https://huggingface.co/datasets/ClareNie/Persona-MME) |
| 📂 **Training Data** | [HF: PersonaVLM-Dataset (80k+ samples)](https://huggingface.co/datasets/ClareNie/PersonaVLM-Dataset) |

## ✒️ Citation
If you find our work helpful, please cite our paper:

```latex
@inproceedings{nie2026personavlm,
  title={PersonaVLM: Long-Term Personalized Multimodal LLMs},
  author={Nie, Chang and Fu, Chaoyou and Zhang, Yifan and Yang, Haihua and Shan, Caifeng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026},
  url={http://arxiv.org/abs/2604.13074}
}