Vero: Open RL Recipe Achieves Strong Visual Reasoning Across Charts, Science, and Spatial Tasks
arXiv·medium signal
Sarch, Cai, Wang et al. release Vero, an open RL recipe for building general visual reasoners that work across charts, science diagrams, spatial understanding, and open-ended tasks. The strongest VLMs show this breadth is achievable, but the training recipe has remained proprietary — Vero opens it up. Includes concrete training methodology with reinforcement learning that practitioners can replicate for their own visual reasoning models.