R1-SyntheticVL: Is Synthetic Data from Generative Models Ready for Multimodal Large Language Model?

If you find this project useful, please give us a star🌟.

Jingyi Zhang^1,2*, Tianyi Lin^3*, Huanjin Yao⁴, Xiang Lan⁵, Shunyu Liu², Jiaxing Huang^1✉️

¹Hong Kong Polytechnic University, ²Nanyang Technological Univeristy, ³Tsinghua University, ⁴ByteDance, ⁵National University of Singapore

^*Equal Contribution, ^✉️Corresponding Author

News

• April 30, 2026. R1-SyntheticVL is accepted by ICML 2026! 🎉
• Feb 03, 2026. We release our paper on arxiv.

Code, model and data will be released soon ...

Main results

We conduct experiments with base model Qwen2.5-VL-7B.

Qualitative illustrations of synthesized samples

Acknowledgment

Our work is built on the following codebases, and we are deeply grateful for their contributions.

• EasyR1
• VLMEvalKit

Citation

We appreciate your citations if you find our paper related and useful to your research!

 @article{zhang2026r1,
  title={R1-SyntheticVL: Is Synthetic Data from Generative Models Ready for Multimodal Large Language Model?},
  author={Zhang, Jingyi and Lin, Tianyi and Yao, Huanjin and Lan, Xiang and Liu, Shunyu and Huang, Jiaxing},
  journal={arXiv preprint arXiv:2602.03300},
  year={2026}
}