A Rubric-based LLM Automatic Grading of Mathematical Reasoning in Self-explained Answers

Abstract

This research addresses challenges in automated math grading by focusing on assessing complex reasoning in high-order math questions through written selfexplanations. We developed a rubric-based scoring system using LLMs, incorporating an algorithmic output checker and self-consistency sampling. Twelve self-explanations were scored, with expert grades as the gold standard. Results show the algorithmic checker outperforms the LLM-based method, and self-consistency sampling enhances alignment with expert judgments. Overall, the approach will offer accurate feedback, reducing teacher workload, boosting student engagement, and enhancing scalability, while indicating a need for automated rubric generation.