VICO: Visual Environments Co-Evolving for Vision-Language Model Reasoning

Reinforcement learning with verifiable rewards (RLVR) has become a standard recipe for post-training vision-language models (VLMs), but it typically assumes a static training environment.