REC-CBM: Rubric-Aware Error-Correction Concept Bottleneck Models for Trustworthy Open-Ended Grading

Researchers propose REC-CBM, a novel machine learning model that combines concept bottleneck models with rubric-aware error correction to automate open-ended educational grading while maintaining transparency and interpretability. Unlike black-box LLM systems, REC-CBM allows educators to verify scoring decisions through human-interpretable concept reasoning, addressing the growing need for trustworthy automated grading in educational settings.

The automation of open-ended grading represents a critical intersection of AI capabilities and educational equity. REC-CBM addresses a fundamental tension in educational technology: the performance advantages of deep learning systems versus the transparency requirements necessary for educator trust and student fairness. While large language models have demonstrated superior grading accuracy, their opaque decision-making processes create friction in adoption, as educators cannot verify why assignments received specific scores.

The research builds on concept bottleneck models, an emerging interpretability paradigm that forces neural networks to make predictions through explicit, human-understandable intermediate representations. REC-CBM extends this framework specifically for grading by incorporating rubric awareness—understanding that grading rubrics have hierarchical, ordered dimensions that standard CBMs fail to capture. The model's error-correction module addresses a practical problem often overlooked in educational AI: human annotators make mistakes when labeling training data, and the system must account for this noise while preserving its interpretability guarantees.

For educational institutions and EdTech developers, REC-CBM offers a pathway to deploy automated grading systems that satisfy regulatory and pedagogical requirements for transparency. Rather than choosing between accuracy and explainability, this approach demonstrates how both can be achieved simultaneously. The work validates practical applicability in realistic classroom settings, suggesting near-term deployment potential.

Future development should focus on scaling across diverse rubric structures, integrating with existing learning management systems, and conducting longitudinal studies measuring educator acceptance and student outcomes when teachers can intervene in automated decisions.

• →REC-CBM combines interpretable concept bottleneck models with rubric-specific design to automate grading while maintaining transparency.
• →The model includes an error-correction module that denoise concept predictions while preserving the explainability guarantees educators require.
• →Experiments demonstrate REC-CBM outperforms black-box LLM baselines while producing more faithful reasoning educators can inspect and verify.
• →The approach enables educators to intervene in and modify automated grading decisions, advancing trust in educational AI systems.
• →The work addresses the critical gap between high-accuracy grading systems and the transparency requirements necessary for institutional adoption.