HY-Wan/Wan-R1

Wan-R1: A Reasoning-via-Video Maze-Solving Model

Fine-tuned on VR-Bench to evaluate and enhance video-based reasoning ability across structured maze environments.

📰 News

• 2025-11-20: Released 5 fine-tuned Wan-R1 models (3D, Regular, Irregular, Sokoban, Trapfield) trained on VR-Bench.
• 2025-12: In-progress: preparing codebase for fine-tuning and evaluation release.

🔧 Future Work

• 📦 Release LoRA fine-tuning scripts based on VR-Bench.
• 📊 Open-source evaluation toolkit for reasoning via video.
• 📁 Provide training logs & hyperparameters for full reproducibility.

🧠 Models

Model	Download	Description
Wan_R1_3d_maze_5B	🤗 HuggingFace	Fine-tuned LoRA for Maze3D tasks (easy, medium, and hard) from the base model Wan2.2-TI2V-5B.
Wan_R1_irregular_maze_5B	🤗 HuggingFace	Fine-tuned LoRA for PathFinder tasks (easy, medium, and hard) from base model Wan2.2-TI2V-5B.
Wan_R1_regular_maze_5B	🤗 HuggingFace	Fine-tuned LoRA for Maze tasks (easy, medium, and hard) from base model Wan2.2-TI2V-5B.
Wan_R1_sokoban_5B	🤗 HuggingFace	Fine-tuned LoRA for Sokoban tasks (easy, medium, and hard) from base model Wan2.2-TI2V-5B.
Wan_R1_trapfield_5B	🤗 HuggingFace	Fine-tuned LoRA for TrapField tasks (easy, medium, and hard) from base model Wan2.2-TI2V-5B.

📑 Citation

If you use this model or the VR-Bench dataset in your work, please cite:

📄 Reasoning via Video: The First Evaluation of Video Models' Reasoning Abilities through Maze-Solving Tasks

@misc{yang2025reasoningvideoevaluationvideo,
      title={Reasoning via Video: The First Evaluation of Video Models' Reasoning Abilities through Maze-Solving Tasks}, 
      author={Cheng Yang and Haiyuan Wan and Yiran Peng and Xin Cheng and Zhaoyang Yu and Jiayi Zhang and Junchi Yu and Xinlei Yu and Xiawu Zheng and Dongzhan Zhou and Chenglin Wu},
      year={2025},
      eprint={2511.15065},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.15065}, 
}