REC-CBM: Rubric-Aware Concept Bottleneck Models for Grading

Researchers have introduced REC-CBM, a rubric-aware error-correction concept bottleneck model aimed at reliable open-ended grading. Traditional concept bottleneck models (CBMs) fall short for this type of assessment as they fail to explicitly represent detailed rubric dimensions, inadequately address the ordinal nature of scoring scales, and overlook reliability concerns in human annotations. By integrating rubric awareness and error correction, REC-CBM seeks to overcome these challenges, promoting transparent and interpretable grading processes. This system is intended to bolster confidence in automated assessments by allowing educators to verify the rationales behind scores. The findings are available on arXiv under the identifier 2605.27402.