Thinking with Programming Vision: Towards a Unified View for Thinking with Images

Overview

• Introduction: A framework leveraging code-as-tool and comprehensive SFT/RL datasets for "thinking with images".

• Features: Supports multi-turn agent loops for the Qwen2.5-VL and Qwen3-VL series.

• Datasets: Includes an SFT dataset constructed using GPT-5-High and an RL dataset covering diverse domains.

Getting Started

Coming soon...