VisualThinker-R1-Zero

Why we included this project

This repo documents a research replication that applies DeepSeek-R1-Zero's reinforcement-learning recipe to a small vision-language model. The authors trained a 2B Qwen2-VL model with GRPO and no supervised fine-tuning, and report that it develops the same self-reflection and lengthened responses seen in the text-only R1-Zero runs, including the kind of 'aha moment' where the model catches and corrects its own mistakes. That makes it a useful reference for researchers and students working on multimodal reasoning, especially anyone curious about how reasoning behavior emerges during RL post-training. A checkpoint is released on Hugging Face, so it also works as a starting point for reproducing or extending the training setup. It is a research artifact rather than a production tool.