HG-Bench: leaderboard

Metric: Step-level micro F1 (0-100), the paper's ranking metric: step-box true and false positives and false negatives pooled over all pages that have gold step boxes; 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 July 2027. 9 models tracked.

Top models

#ModelScore
1Gemini 3 Pro (Preview)48.22
2GLM-5V Turbo26.29
3Qwen 3.5 397B A17B18.15
4Kimi K2.57.42
5Claude Sonnet 4.61.63
6GPT-5.41.55

Interactive version: theaggregate.ai/benchmark?slug=hg-bench · How It Works · Data refreshed daily, snapshot 2026-09-29.