VisualThinker-R1-Zero applies GRPO-based reinforcement learning to train Qwen2-VL-2B on visual reasoning tasks without supervised fine-tuning or reward models. The project demonstrates emergent self-reflection and correction behaviors in visual reasoning, successfully reproducing the ‘aha moment’ and increasing response length observed in DeepSeek-R1-Zero. This enables reasoning capabilities to emerge from pure RL training on vision-centric tasks.