This repository implements a lightweight research framework for improving vision-language model reasoning with reinforcement learning. The project focuses on Qwen2.5-VL and studies how PPO, GRPO, ...