We propose a retrieval-augmented method for grading short-answer exam responses that grounds an LLM grader in a rubric and a bank of exemplar answers, improving agreement with human markers while making each grade explainable.
Read full text
1. Introduction
Automated grading of open-ended responses is hard because teh space of acceptable answers is large. Prior work (Ahmed, 2021) shows that rubric-grounded models agree better with human markers.
2. Method
We retrieve exemplar answers for each question and condition the grader on them. We present a compact transformer architecture that achieves competitive accuracy on low-resource text-classification tasks while reducing parameter count.
3. Results
Across three courses the method improves quadratic-weighted kappa with human graders by eleven points and produces a short rationale for every grade.
4. Conclusion
Retrieval-augmented grading is accurate and explainable, and is practical for classroom use.