VLM-R1

Why we included this project

VLM-R1 shows what happens when you apply the R1-style reinforcement learning recipe to a vision-language model, and it publishes everything needed to follow along: training code, the exact datasets, and pretrained checkpoints. Teams fine-tuning a multimodal model with GRPO often find practical details spread thin across papers; here it is built on Qwen2.5-VL, with reproducible scripts and the data used. The released checkpoints cover referring expression comprehension, open-vocabulary detection, and visual math reasoning, so you can either reuse them for those jobs or study how the method was trained. The README also documents a useful finding: reinforcement learning generalizes better to out-of-domain data than supervised fine-tuning as training steps increase. With live demos and a tech report, it is a concrete starting point for experimenting with RL in multimodal settings.