Bounding Boxes to Improve Small Language Model Performance on Vision-Based Grading Tasks

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Lachlan McGinness
• Venue: arXiv: cs.CV
• Published: 2026-07-21
• Collected: 2026-07-22
• URL: http://arxiv.org/abs/2607.18767v1
• PDF: https://arxiv.org/pdf/2607.18767v1
• DOI: UNKNOWN

Plain-Language Brief

A research paper proposes using bounding boxes to improve the accuracy and efficiency of Small Language Models (SLMs) when grading handwritten student exams. The study finds that this method significantly enhances performance across various models, making it a valuable pre-processing step for large-scale educational assessments.

What the Document Says

• The paper discusses the use of Small Language Models in education due to their privacy, cost, and scalability benefits.
• SLMs struggle with complex vision-based tasks like grading handwritten student exams due to high computational costs and visual distractions.
• The study investigates if cropping student responses using bounding boxes can improve SLM performance on a short-answer grading task.
• The research is based on a dataset from the 2025 Australian Physics Olympiad.
• Several models, ranging from 4B to 72B parameters, were evaluated under varying Chain of Thought prompting and image cropping conditions.
• Results show that using bounding boxes increases grading accuracy and reduces computational costs (FLOPs) across models.
• The authors conclude that bounding boxes are essential for deploying SLMs in large-scale, vision-based educational assessments.

Key Concepts Mentioned

• Small Language Models
• bounding boxes
• grading tasks
• computational cost
• FLOPs

Claims Made by the Document

• Cropping student responses using bounding boxes improves the accuracy of SLMs on a short-answer grading task.
• Using bounding boxes reduces computational cost (FLOPs) across models.

Evidence Used

• Results from evaluating several models under varying conditions

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

The deployment of Small Language Models (SLMs) in educational settings offers significant advantages in terms of privacy, cost, and scalability. However, SLMs often struggle with complex vision-based tasks, such as grading handwritten student exams, due to the high computational cost of processing large images and the visual distractions present on a full page. In this paper, we investigate whether cropping student responses using bounding boxes can improve the accuracy and computational efficiency of SLMs on a short-answer grading task. Using a dataset of scanned handwritten responses from the 2025 Australian Physics Olympiad, we evaluate the performance of several models ranging from 4B to 72B parameters under varying conditions of Chain of Thought (CoT) prompting and image cropping. Our results demonstrate that using bounding boxes significantly improves grading accuracy and reduces computational cost (FLOPs) across models. We conclude that bounding boxes are a crucial pre-processing step for deploying SLMs in large-scale, vision-based educational assessments.