Confidence Estimation in Automatic Short Answer Grading with LLMs
Researchers combine model signals and data uncertainty for safer AI grading.
A new study tackles a critical gap in AI-assisted education: LLMs like GPT-4 can grade short answers without fine-tuning, but their grading isn't perfect. To decide when to trust the AI vs. escalate to a human, reliable confidence estimates are essential. The researchers from multiple institutions systematically compared three model-based confidence methods: having the LLM verbalize its confidence, using latent representations, and analyzing consistency across repeated outputs. They found that none alone reliably captured uncertainty.
To fix this, they introduced a hybrid framework that augments model-based signals with dataset-derived aleatoric uncertainty—the inherent ambiguity in student responses. They cluster semantically similar answers via embeddings, then measure within-cluster grading variability. This hybrid approach yielded more accurate confidence estimates and significantly improved selective grading performance (knowing when to defer to a human). The work paves the way for trustworthy AI assessment systems that know their limits.
- Compared three confidence estimation strategies: verbalized, latent, and consistency-based
- Proposed hybrid framework that adds dataset-derived aleatoric uncertainty via clustering student responses
- Hybrid method outperforms single-source approaches in selective grading tasks
Why It Matters
Makes AI grading systems more trustworthy by knowing when to defer to human judges.