Confidence-Aware Automated Assessment of Student-Drawn Scientific Models

This article presents a vision-based automated scoring system for student-generated scientific drawings, utilizing a Vision Transformer (ViT) with parameter-efficient adaptation. The proposed confidence-aware scoring framework assesses response-level confidence from predictive distributions, allowing for automated scoring of high-confidence responses while deferring uncertain cases for human evaluation. This approach enhances scoring reliability and supports a balance between automation and assessment accuracy, which is crucial for scalable educational applications aligned with the Next Generation Science Standards (NGSS).