PixelEyes enhances active visual search in MLLMs by delegating fine-grained localization to a specialized perception tool, thereby achieving efficient and accurate multi-turn visual reasoning.
This repository contains the weights for PixelEyes, a multi-turn visual reasoning agent that explicitly decouples reasoning from perception, introduced in the paper PixelEyes: Decoupling Perception and Reasoning for Pinpoint Visual Evidence Seeking.
For more details on environment setup, training, and evaluation, please visit the GitHub repository.
Citation
If you find this project helpful in your research, please cite our paper:
@misc{gong2026pixeleyesdecouplingperceptionreasoning,
title={PixelEyes: Decoupling Perception and Reasoning for Pinpoint Visual Evidence Seeking},
author={Dengxian Gong and Yuanzheng Wu and Haobo Yuan and Zhengdong Hu and Tao Zhang and Yikang Zhou and Shihao Chen and Quanzhu Niu and Kai Wang and Jason Li and Haochen Wang and Lu Qi and Shunping Ji and Ming-Hsuan Yang},
year={2026},
eprint={2607.00115},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2607.00115},
}