VisPlay

VisPlay: Self-Evolving Vision-Language Models from Images

Yicheng He, Chengsong Huang, Zongxia Li et al. · CVPR 2026

Read paper · arXiv:2511.15661

Build a visual curriculum from unlabeled images.

What you will implement

VisPlay Image-Grouped Rewards

• 111. VisPlay Image-Grouped Rewards

  VisPlay Image-Grouped Rewards

  Compute per-image repetition and combine uncertainty with format gating.