The era of Large Reasoning Models (LRMs) has begun, bringing new opportunities and challenges to the computer vision and multimodal AI community. While Large Language Models (LLMs) excel at semantic understanding, the next frontier is System-2 style slow thinking—reasoning beyond pattern recognition to multi-step, causal, and neuro-symbolic reasoning.

MARS2 (Multimodal Reasoning and Slow Thinking) is an official workshop of ICCV 2025. Our goal is to unite researchers from computer vision, multimodal learning, and reasoning to explore how AI systems can achieve flexible, robust, and interpretable reasoning.

• Workshop at ICCV 2025
  Hosted in Honolulu, Hawaii, featuring keynote talks from world-renowned researchers in AI and computer vision.

• High-Stakes Reasoning Competition
  A large-scale challenge with a ¥100,000 (~$14,000) prize pool, testing MLLMs on a diverse set of reasoning-oriented benchmarks.

• Open-Source Baselines & Experiments
  Provides official repositories with baseline implementations using state-of-the-art models like Qwen-VL, InternVL, and Ferret to facilitate research and participation.

• Fostering Community Collaboration
  Bringing together experts to define the next frontier of AI by bridging computer vision, NLP, and System-2 reasoning.

The workshop hosts the MARS2 Multimodal Reasoning Challenge, designed to push the boundaries of current MLLMs on complex reasoning tasks. The challenge is divided into distinct tracks, each with corresponding baseline models and official code repositories.

The challenge will test performance on benchmarks including LENS, AdsVQA, open-ended reasoning tasks, designed for System-2 evaluation.

This organization serves as the official hub for all workshop activities, competition tracks, and related experiments.

⚠️ These model-specific folders provide experimental support for benchmarking and challenge tasks. They are intended as baselines and resources for participants.

We gratefully acknowledge the contributions of the following open-source projects, which form the foundation for our experimental extensions and benchmarking.

• Qwen: A powerful series of large language and vision-language models by Alibaba Cloud.
• InternVL: A foundational open-source vision-language model designed for advanced multimodal understanding.
• Ferret: An MLLM capable of referring and grounding anything anywhere at any granularity.
• Groma: An MLLM with exceptional region understanding and visual grounding capabilities.

If you use MARS2 materials, benchmarks, or code in your research, please cite our workshop:

```
@inproceedings{xu2025mars2,
author    = {Xu, Peng and Xiong, Shengwu and Zhang, Jiajun and Chen, Yaxiong and Zhou, Bowen and Loy, Chen Change and Clifton, David and Lee, Kyoung Mu and Van Gool, Luc and others},
title     = {MARS2 2025 Challenge on Multimodal Reasoning: Datasets, Methods, Results, Discussion, and Outlook},
booktitle = {ICCV Workshop},
year      = {2025}
}
```

```
@article{yao2025lens,
title={LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models},
author={Yao, Ruilin and Zhang, Bo and Huang, Jirui and Long, Xinwei and Zhang, Yifang and Zou, Tianyu and Wu, Yufei and Su, Shichao and Xu, Yifan and Zeng, Wenxi and Yang, Zhaoyu and Li, Guoyou and Zhang, Shilan and Li, Zichan and Chen, Yaxiong and Xiong, Shengwu and Xu, Peng and Zhang, Jiajun and Zhou, Bowen and Clifton, David and Van Gool, Luc},
journal={arXiv preprint arXiv:2505.15616},
year={2025}
}
```

```
@inproceedings{long2025adsqa,
author    = {Long, Xinwei and Tian, Kai and Xu, Peng and Jia, Guoli and Li, Jingxuan and Yang, Sa and Shao, Yihua and Zhang, Kaiyan and Jiang, Che and Xu, Hao and Liu, Yang and Ma, Jiaheng and Zhou, Bowen},
title     = {AdsQA: Towards Advertisement Video Understanding},
booktitle = {ICCV},
year      = {2025}
}
```