csfufu/Revisual-R1-Coldstart

🌟 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. Text Cold-Start — seed deep reflection
  2. Multimodal RL — align vision & logic
  3. Text RL — polish fluency & brevity

• PAD (Prioritized Advantage Distillation) keeps gradients alive.

• Efficient-Length Reward = concise, self-reflective CoT.

---

📚 Resources

• Paper
• Code

---

📌 Citation

@misc{chen2025advancingmultimodalreasoningoptimized,
  title         = {Advancing Multimodal Reasoning: From Optimized Cold Start to Staged Reinforcement Learning},
  author        = {Shuang Chen and Yue Guo and Zhaochen Su and Yafu Li and Yulun Wu and Jiacheng Chen and
                   Jiayu Chen and Weijie Wang and Xiaoye Qu and Yu Cheng},
  year          = {2025},
  eprint        = {2506.04207},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2506.04207}
}

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