Benchmark RadarHakushoBench
HakushoBench is a Japanese chart and table VQA benchmark built from 33 governmental white papers, containing 2,053 images across over 10 image types with manually annotated QA pairs. It evaluates vision-language models on deep holistic understanding of charts and tables.
- Released
- 2026-05-31
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
HakushoBench addresses the scarcity of non-English benchmarks for chart and table understanding, providing a challenging evaluation for VLMs in Japanese document domains. It reveals a notable performance gap between open-weight and proprietary models, highlighting areas for improvement in multilingual document AI.
Motivation
Understanding chart and table images is essential for applying vision-language models (VLMs) to real-world document understanding.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.