MVToolBench: Benchmarking Multiple Tool Usage

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

• Paper: https://arxiv.org/abs/2512.03746
• Code: https://github.com/ByteDance-BandAI/CodeVision

Dataset Details

An example:

extra_info:
{
'answer': 'park',
'bbox': [1060, 82, 1072, 286],
'dataset_name': 'hiertext',
'index': 0,
'question': 'What is the shortest word on the line that starts with the text "LANDSCAPE..."?',
'split': 'train',
'transform': ['crop', 'rotate_90']
}

prompt:
[{
'content': '<image>\nWhat is the shortest word on the line that starts with the text "LANDSCAPE..."?',
'role': 'user'
}]

The red box (in extra_info bbox) is the answer region.

Citation

If you find the dataset useful, please cite the paper:

@article{guo2025thinking,
  title={Thinking with Programming Vision: Towards a Unified View for Thinking with Images},
  author={Guo, Zirun and Hong, Minjie and Zhang, Feng and Jia, Kai and Jin, Tao},
  journal={arXiv preprint arXiv:2512.03746},
  year={2025}
}