MARS2 2025 多模态推理挑战：数据集、方法、结果、讨论与展望
计算机视觉与模式识别
2025-09-18 v1
摘要
本文综述了 MARS2 2025 多模态推理挑战。我们旨在通过一个大型基准测试，将多模态机器学习和大语言模型（LLMs）中的不同方法汇聚。我们希望更好地帮助研究人员跟随这一非常动态的领域的最新进展。同时，日益增长的测试平台正在推动通用大型语言模型的演进。因此，今年的 MARS2 持续聚焦现实场景和专业场景，以拓展多模态大语言模型（MLLMs）的多模态推理应用。我们的组织团队发布了两个量身定制的数据集 Lens 和 AdsQA 作为测试集，分别支持对 12 种日常场景中的通用推理以及广告视频中的领域特定推理。我们评估了 40 多种基线方法，包括通用型 MLLMs 和任务特定模型，并开放了三个竞赛轨道，即现实场景下的视觉定位（VG-RS）、具备空间感知能力的视觉问答（VQA-SA）以及创意广告视频中的视觉推理（VR-Ads）。最终，76 支来自知名学术和工业机构的团队报名参赛，其中 40 多支有效提交（总计 1200+ 份）已进入我们的排名列表。我们的数据集、代码集（包含 40 多种基线方法和 15 多种参赛队方法）以及排名均公开在 MARS2 工作坊网站及我们的 GitHub 组织页面 https://github.com/mars2workshop/ 上，其中将持续提供我们的更新和后续活动公告。
引用
@article{arxiv.2509.14142,
title = {MARS2 2025 Challenge on Multimodal Reasoning: Datasets, Methods, Results, Discussion, and Outlook},
author = {Peng Xu and Shengwu Xiong and Jiajun Zhang and Yaxiong Chen and Bowen Zhou and Chen Change Loy and David A. Clifton and Kyoung Mu Lee and Luc Van Gool and Ruiming He and Ruilin Yao and Xinwei Long and Jirui Huang and Kai Tian and Sa Yang and Yihua Shao and Jin Feng and Yue Zhong and Jiakai Zhou and Cheng Tang and Tianyu Zou and Yifang Zhang and Junming Liang and Guoyou Li and Zhaoxiang Wang and Qiang Zhou and Yichen Zhao and Shili Xiong and Hyeongjin Nam and Jaerin Lee and Jaeyoung Chung and JoonKyu Park and Junghun Oh and Kanggeon Lee and Wooseok Lee and Juneyoung Ro and Turghun Osman and Can Hu and Chaoyang Liao and Cheng Chen and Chengcheng Han and Chenhao Qiu and Chong Peng and Cong Xu and Dailin Li and Feiyu Wang and Feng Gao and Guibo Zhu and Guopeng Tang and Haibo Lu and Han Fang and Han Qi and Hanxiao Wu and Haobo Cheng and Hongbo Sun and Hongyao Chen and Huayong Hu and Hui Li and Jiaheng Ma and Jiang Yu and Jianing Wang and Jie Yang and Jing He and Jinglin Zhou and Jingxuan Li and Josef Kittler and Lihao Zheng and Linnan Zhao and Mengxi Jia and Muyang Yan and Nguyen Thanh Thien and Pu Luo and Qi Li and Shien Song and Shijie Dong and Shuai Shao and Shutao Li and Taofeng Xue and Tianyang Xu and Tianyi Gao and Tingting Li and Wei Zhang and Weiyang Su and Xiaodong Dong and Xiao-Jun Wu and Xiaopeng Zhou and Xin Chen and Xin Wei and Xinyi You and Xudong Kang and Xujie Zhou and Xusheng Liu and Yanan Wang and Yanbin Huang and Yang Liu and Yang Yang and Yanglin Deng and Yashu Kang and Ye Yuan and Yi Wen and Yicen Tian and Yilin Tao and Yin Tang and Yipeng Lin and Yiqing Wang and Yiting Xi and Yongkang Yu and Yumei Li and Yuxin Qin and Yuying Chen and Yuzhe Cen and Zhaofan Zou and Zhaohong Liu and Zhehao Shen and Zhenglin Du and Zhengyang Li and Zhenni Huang and Zhenwei Shao and Zhilong Song and Zhiyong Feng and Zhiyu Wang and Zhou Yu and Ziang Li and Zihan Zhai and Zijian Zhang and Ziyang Peng and Ziyun Xiao and Zongshu Li},
journal= {arXiv preprint arXiv:2509.14142},
year = {2025}
}
备注
ICCV 2025 MARS2 Workshop and Challenge "Multimodal Reasoning and Slow Thinking in the Large Model Era: Towards System 2 and Beyond''