A Two-Stage LLM Pipeline for Handwritten Mathematics Autograding

While question-asking platforms have provided a wide array of pedagogical benefits and have decreased grader workload in university classes, they are limited in what types of inputs are accepted. Recent work has explored using Large Language Models to expand what can be automatically graded. In this study, we examine their use for grading handwritten mathematics questions via rubrics. Although much work remains, our preliminary results suggest that using separate prompts to first extract text and then evaluate rubric items enables LLMs to distinguish fully correct solutions from those requiring further review, thereby reducing grader workload.