LLMs outperform supervised models for cross-dataset Bloom classification

Bloom's taxonomy classification of assessment questions helps reduce instructor workload, but labeling is subjective and prior machine learning and deep learning approaches were rarely tested across datasets—leaving real-world generalizability unclear. Researchers Abdolali Faraji, Mohammadreza Molavi, and colleagues systematically evaluated cross-dataset generalization of existing supervised ML/DL methods and assessed LLMs with multiple prompting strategies on five datasets. The best LLM prompting strategy combined in-context examples with course-specific action verbs.

Supervised ML/DL models degraded substantially on unseen datasets, while LLMs remained more stable, suggesting a robust alternative for diverse educational contexts. Based on the best prompting strategy, the team built a lightweight user interface that automatically classifies large question banks; a usability study indicated low workload and high usability. This work highlights the advantage of LLMs for real-world educational deployment where training data often differs from test data.