MiMo-VL 技术报告
计算与语言
2025-06-05 v1
摘要
我们开源 MiMo-VL-7B-SFT 和 MiMo-VL-7B-RL 两个强大的视觉语言模型，分别在通用视觉理解和多模态推理方面均实现最先进的性能。MiMo-VL-7B-RL 在 40 个评估任务中超过 Qwen2.5-VL-7B，得分 59.4，超过最高可达 78B 参数的模型。在 GUI grounding 应用中，达成 56.1 的 OSWorld-G 分数，甚至超过了专门模型如 UI-TARS。我们的训练结合四阶段预训练 (2.4 万亿 token) 与混合 on-policy 强化学习 (MORL) 集成多样化奖励信号。我们识别出在预训练阶段纳入高质量推理数据与长链式思维的重要性，并在面对同时多域优化挑战的情况下，混合 RL 的优势。我们还贡献了一个覆盖 50+ 任务的全面评估套件，以促进可重复性并推动该领域发展。模型检查点和完整评估套件已公开于 https://github.com/XiaomiMiMo/MiMo-VL。
引用
@article{arxiv.2506.03569,
title = {MiMo-VL Technical Report},
author = {Core Team and Zihao Yue and Zhenru Lin and Yifan Song and Weikun Wang and Shuhuai Ren and Shuhao Gu and Shicheng Li and Peidian Li and Liang Zhao and Lei Li and Kainan Bao and Hao Tian and Hailin Zhang and Gang Wang and Dawei Zhu and Cici and Chenhong He and Bowen Ye and Bowen Shen and Zihan Zhang and Zihan Jiang and Zhixian Zheng and Zhichao Song and Zhenbo Luo and Yue Yu and Yudong Wang and Yuanyuan Tian and Yu Tu and Yihan Yan and Yi Huang and Xu Wang and Xinzhe Xu and Xingchen Song and Xing Zhang and Xing Yong and Xin Zhang and Xiangwei Deng and Wenyu Yang and Wenhan Ma and Weiwei Lv and Weiji Zhuang and Wei Liu and Sirui Deng and Shuo Liu and Shimao Chen and Shihua Yu and Shaohui Liu and Shande Wang and Rui Ma and Qiantong Wang and Peng Wang and Nuo Chen and Menghang Zhu and Kangyang Zhou and Kang Zhou and Kai Fang and Jun Shi and Jinhao Dong and Jiebao Xiao and Jiaming Xu and Huaqiu Liu and Hongshen Xu and Heng Qu and Haochen Zhao and Hanglong Lv and Guoan Wang and Duo Zhang and Dong Zhang and Di Zhang and Chong Ma and Chang Liu and Can Cai and Bingquan Xia},
journal= {arXiv preprint arXiv:2506.03569},
year = {2025}
}
备注
32 pages