OpenThinkIMG: Learning to Think with Images via Visual Tool Reinforcement Learning

Zhaochen Su, Linjie Li, Mingyang Song, Yunzhuo Hao, Zhengyuan Yang, Jun Zhang, Guanjie Chen, Jiawei Gu, Juntao Li, Xiaoye Qu, Yu Cheng

arXiv preprint arXiv:2505.08617

Abstract

OpenThinkIMG is an open-source framework for tool-augmented large vision-language models. It provides standardized visual tool interfaces, scalable trajectory generation, and V-ToolRL, a reinforcement learning method that enables models to learn adaptive policies for invoking external visual tools during reasoning.

Paper  |  Code