HG-Bench - Answer-Region F1: leaderboard
HG-Bench (2026) tests page-aware answer-region grounding in multi-page handwritten K-12 homework: over 500 human-annotated samples, a vision-language model must box each complete answer region and its ordered step regions. This board is the answer-region F1 scaled to 0-100, averaged per page and sample over the samples whose output parsed; higher is better.
Metric: Answer-region F1 (0-100): per-page box F1 for complete answer regions, averaged within each sample and then over the successfully parsed samples; 500 human-annotated multi-page handwritten K-12 homework samples; the model outputs page-aware question-level answer boxes and ordered step boxes as JSON with normalized 0-1000 coordinates, one shared prompt, default decoding, one format-reminder retry; higher is better. Source: arxiv.org. Saturation forecast: Around October 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 50.9 | #113 |
| 2 | GLM-5V Turbo | 46.69 | #344 |
| 3 | Qwen 3.5 397B A17B | 42.71 | #246 |
| 4 | Kimi K2.5 | 31.21 | #256 |
| 5 | Claude Sonnet 4.6 | 16.83 | #161 |
| 6 | GPT-5.4 | 14.91 | #123 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=hg-bench-answer-region-f1 · How It Works · Data refreshed daily, snapshot 2026-11-13.