HG-Bench: A Benchmark for Multi-Page Handwritten Answer-Region Grounding in Automated Homework Assessment

2026-06-24

AI Summary

This study addresses the lack of evaluation mechanisms for precisely localizing multi-page handwritten solutions and their reasoning steps in existing automated grading approaches. We propose the first two-level region localization benchmark tailored for multi-page handwritten responses, which requires models to simultaneously identify the full answer region across a sequence of pages and its constituent, ordered step subregions, accompanied by a page-aware evaluation protocol. To support this benchmark, we construct a large-scale, manually annotated dataset featuring hierarchical annotations at both the problem and step levels—the first of its kind—and introduce a dual-metric evaluation framework based on Full Answer (FA) and Fine-grained Step (FSm) scores. Experiments reveal that zero-shot models exhibit limited performance (FA/FSm ≤ 55.22%/48.22%), whereas a fine-tuned GLM-4.6V 9B model achieves substantial gains (74.97%/72.26%), highlighting a critical capability gap in current models for fine-grained step localization.

Abstract

Automated homework assessment depends not only on recognizing student answers, but also on accurately locating where each answer and each intermediate reasoning step appears in noisy, multi-page handwritten work. This paper addresses the missing evaluation setting of page-aware, two-level answer-region grounding: given a sequence of homework page images, a model must localize complete answer regions and their ordered step-level subregions. We introduce HG-Bench, a benchmark of 500 human-annotated K-12 homework samples curated from a 1,489,278-image source pool, with question-level and step-level boxes linked by a hierarchical containment constraint. HG-Bench is paired with a page-aware evaluation protocol that separately measures complete-answer localization (FA) and step-level decomposition (FSm), revealing whether models truly ground the spatial structure of student reasoning rather than merely parse visible text. Across frontier closed-source APIs and competitive open-weight VLMs, no zero-shot system exceeds 55.22% on FA or 48.22% on FSm, while a GLM-4.6V 9B reference model fine-tuned on ~10k in-domain examples reaches 74.97/72.26. These results identify step-level handwritten grounding as a concrete capability gap and provide a reproducible benchmark, evaluation protocol, and trained reference point for future work on automated homework assessment.

Problem

• handwritten answer grounding
• multi-page localization
• step-level reasoning
• automated homework assessment
• spatial structure grounding

Innovation

• answer-region grounding
• handwritten homework assessment
• multi-page VLM benchmark
• step-level reasoning localization
• hierarchical spatial annotation