csfufu/Revisual-R1-final

🌟 ReVisual-R1 (7B) — Open-Source Multimodal Reasoner

One cold-start, two RL stages, endless reasoning power.

🔑 Highlights

• —SOTA on 9 tough benchmarks covering visual–math + text reasoning.
• —Three-Stage SRO Training

1. 1.Text Cold-Start — seed deep reflection
2. 2.Multimodal RL — align vision & logic
3. 3.Text RL — polish fluency & brevity
4. 4.PAD (Prioritized Advantage Distillation) keeps gradients alive.
5. 5.Efficient-Length Reward = concise, self-reflective CoT.

📚 Resources

• —Paper
• —Code

📌 Citation

@article{chen2025advancing,
  title={Advancing Multimodal Reasoning: From Optimized Cold Start to Staged Reinforcement Learning},
  author={Chen, Shuang and Guo, Yue and Su, Zhaochen and Li, Yafu and Wu, Yulun and Chen, Jiacheng and Chen, Jiayu and Wang, Weijie and Qu, Xiaoye and Cheng, Yu},
  journal={arXiv preprint arXiv:2506.04207},
  year={2025}
}

Take ReVisual-R1 for a spin and let us know what you build! 🎯