HakushoBench: A Japanese Chart and Table VQA Benchmark from Governmental White Papers

Researchers introduced HakushoBench, a Japanese chart-and-table VQA benchmark built from 33 government white papers.

The dataset includes 2,053 chart and table images with manually annotated questions and answers across more than 10 image types.

It is designed to test deeper understanding of visual information in Japanese documents, not just surface recognition.

Results show open-weight vision-language models struggle, with the best reaching 58.6% accuracy and proprietary models leading by 34.9 points.