HakushoBench (CoT): leaderboard
HakushoBench (Sugiura et al., LLM-jp, 2026) asks vision-language models short-answer questions about 2,053 Japanese chart and table images from 33 government white papers, written to need whole-image, multi-hop, counting or background-knowledge reasoning. This board uses the chain-of-thought prompt; accuracy in percent, judged by GPT-5.1 and averaged over three runs, higher is better.
Metric: Accuracy (%), chain-of-thought prompt (think step by step, then answer); 2,053 manually written short-answer questions on chart and table images from 33 Japanese government white papers, judged correct or incorrect by GPT-5.1, mean of three runs, temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 (Medium) | 67.9 | #229 (GPT-5.1) |
| 2 | GPT-4o (2024-11-20) | 54.1 | #866 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=hakushobench-cot · How It Works · Data refreshed daily, snapshot 2026-11-13.