Confidence-Aware AI Scores Student Drawings for Science Education

Student-generated drawings are a valuable tool in science education for assessing conceptual understanding, particularly in tasks aligned with Next Generation Science Standards (NGSS). However, interpreting and scoring these complex visual representations typically requires expert human judgment, making large-scale assessment both costly and difficult to sustain in classroom settings. To address this challenge, new research explores automated scoring of student scientific drawings using a vision-based model. The study evaluates a Vision Transformer (ViT) with parameter-efficient adaptation and introduces a novel confidence-aware scoring framework. This framework derives response-level confidence from the model's predictive distributions during test time. The key innovation of this confidence-aware approach is its ability to enable selective automation. High-confidence responses can be scored automatically, significantly reducing the human workload, while uncertain cases are flagged and deferred for expert human review. Experiments conducted on six NGSS-aligned middle school assessment items demonstrated that this method not only improves scoring reliability but also supports a practical trade-off between automated coverage and the risk of mis-scoring, highlighting its value for trustworthy educational assessment.