Confidence Estimation in Automatic Short Answer Grading with LLMs

AI grading systems need multiple confidence signals to be trustworthy: The paper found that relying on a single confidence measure from LLMs is insufficient for educational grading. Instead, combining model-based confidence (what the AI "thinks" about its answer) with dataset-derived uncertainty (patterns in how students actually answer) produces more reliable results for determining when human teachers should review AI grades.

Current LLM confidence signals can be misleading on their own: Three different methods tested for extracting confidence from LLMs—verbalizing uncertainty, using internal model signals, and checking consistency across multiple attempts—all fell short individually. This suggests that educators cannot simply ask an AI "how confident are you?" and trust the answer without additional verification methods.

Student response patterns reveal important grading uncertainties: The research identified that clustering similar student answers and measuring disagreement within those clusters helps identify genuinely ambiguous responses that deserve human review. This dataset-derived uncertainty captures real-world grading difficulty that the AI model alone might miss.

Hybrid confidence framework improves human-AI collaboration in education: Combining multiple uncertainty sources improved the system's ability to flag uncertain grades for teacher review, supporting safer "human-in-the-loop" assessment where AI handles routine cases while humans focus on borderline or complex answers.

This work addresses a critical gap in AI-assisted education: By tackling confidence estimation specifically, the paper helps make LLM-based grading systems more transparent and accountable—essential for educational contexts where grading decisions directly impact student outcomes.