Vero-600k is a multimodal, multi-task visual reinforcement learning dataset released by Princeton University's zlab team in 2026, accompanied by the research paper "Vero: Scaling Multi-Task Visual Reinforcement Learning", designed for multi-task visual reasoning reinforcement learning training and evaluation of vision-language models.

The dataset integrates 59 sub-datasets, covering six visual reasoning categories: STEM reasoning, charts and OCR, spatial reasoning and action, knowledge and recognition, visual grounding and counting, and image description and instruction following, totaling 600,000 multimodal reinforcement learning image-text samples.

Citation

@article{sarch2026vero,
  title = {Vero: An Open RL Recipe for General Visual Reasoning},
  author = {Sarch, Gabriel and Cai, Linrong and Wang, Qunzhong and Wu, Haoyang and Danqi Chen and Zhuang Liu},
  year = {2026},
  journal = {arXiv preprint arXiv:2604.04917},
}