HG-Bench: A Benchmark for Multi-Page Handwritten Answer-Region Grounding in Automated Homework Assessment
Researchers introduce HG-Bench, a benchmark dataset of 500 annotated homework samples for evaluating automated grading systems' ability to locate and decompose handwritten student answers across multiple pages. Current AI models, including frontier VLMs, achieve less than 55% accuracy on complete answer localization, revealing a significant capability gap in understanding spatial reasoning structures in handwritten documents.
HG-Bench addresses a critical blind spot in automated education technology: while OCR and text recognition have matured, understanding the spatial organization of student reasoning in handwritten work remains largely unsolved. The benchmark moves beyond simple answer extraction to require models to hierarchically ground both complete answer regions and intermediate step-level decompositions—a task that mirrors how human graders mentally parse messy, multi-page submissions. This represents a shift from text-centric evaluation toward spatial-reasoning understanding.
Automated homework assessment sits at the intersection of educational technology and AI capability assessment. Existing solutions focus on recognizing correct answers rather than validating student reasoning pathways, limiting their utility for formative assessment and learning analytics. HG-Bench's hierarchical annotation structure and page-aware evaluation protocol establish reproducible benchmarks where frontier closed-source APIs and competitive open-weight models all underperform, with zero-shot systems capping around 55% on complete answer localization.
The performance gap has immediate implications for EdTech developers and institutions considering automated grading deployment. Current models cannot reliably understand multi-step mathematical proofs, scientific explanations, or other reasoning-dependent work at the K-12 level, indicating that widespread rollout remains premature. The reference model's 75% accuracy after fine-tuning on 10,000 examples suggests domain-specific training can close gaps, but annotation requirements may prove prohibitive at scale.
Future work must address the disconnect between visual understanding and logical structure parsing. Success requires models that understand both document layout and mathematical/scientific reasoning—skills rarely developed in general-purpose VLMs trained on internet data. This benchmark crystallizes a concrete technical challenge that educational AI must solve before deployment.
- →Current frontier AI models fail to localize handwritten student answers at scale, achieving under 55% on complete answer regions
- →Step-level reasoning decomposition proves significantly harder than complete answer localization, exposing a spatial-reasoning understanding gap
- →Fine-tuning on domain-specific homework data improves performance to 75%, but requires substantial annotation effort
- →HG-Bench provides the first reproducible benchmark for page-aware multi-level grounding in educational assessment contexts
- →Automated homework grading systems are not production-ready for reasoning-dependent assignments in K-12 education