VisDoTQA: Enhancing Visual Reasoning through Human-Like Interpretation Grounding and Decomposition of Thought

This repository releases VisDoTQA, the public benchmark introduced in our paper VisDoT : Enhancing Visual Reasoning through Human-Like Interpretation Grounding and Decomposition of Thought.

The canonical publication record is the ACL Anthology page for Findings of the Association for Computational Linguistics: EACL 2026, and an additional mirror is available on arXiv:2603.11631.
See our GitHub repository for the synchronized source release and updates.
See the paper on ACL Anthology, on arXiv:2603.11631, or via DOI.

Highlights

• We release VisDoTQA, a public benchmark for evaluating visual grounding and compositional reasoning on chart images.
• The benchmark contains 1,120 QA pairs built from 609 held-out charts.
• VisDoTQA covers four perceptual task families: Position, Length, Pattern, and Extract.
• This Hugging Face repository releases the public benchmark test split only. The full research dataset described in the paper contains 331,969 QA pairs and is not included here.

Dataset Structure

• Split: test
• Images: test/images/
• Metadata: test/metadata.jsonl

Each example contains:

• file_name: relative path to the chart image
• imgname: image filename
• query: benchmark question
• label: ground-truth answer
• source: VisDoTQA task category (Position, Length, Pattern, Extract)

Links

• Paper (ACL Anthology, canonical)
• Paper (arXiv:2603.11631 mirror)
• DOI
• GitHub Repository

Contact

If you have questions about this dataset release, please use the GitHub repository.

Citation

@inproceedings{lee2026visdot,
  title={VisDoT : Enhancing Visual Reasoning through Human-Like Interpretation Grounding and Decomposition of Thought},
  author={Lee, Eunsoo and Lee, Jeongwoo and Hong, Minki and Choi, Jangho and Kim, Jihie},
  booktitle={Findings of the Association for Computational Linguistics: EACL 2026},
  pages={610--640},
  year={2026},
  doi={10.18653/v1/2026.findings-eacl.30},
  url={https://aclanthology.org/2026.findings-eacl.30/}
}