On this page

Generating In-Context, Personalized Feedback for Intelligent Tutors with Large Language Models — Reddig, Arora & MacLellan (2025) run three studies evaluating whether GPT-4 can diagnose student errors and generate corrective feedback within an Intelligent Tutoring system, the Apprentice Tutor College Algebra platform. Across 6,926 logged transactions, GPT-4 diagnosed student errors with 87.8% accuracy when given rich problem context, though accuracy fell sharply for multi-step problems and responses containing more than one error. The authors find that roughly a third of generated hints were too general, incorrect, or prematurely gave away the answer, and that LLM-based automated evaluation methods for feedback quality were misaligned with human judgment — signaling that LLM-generated Feedback requires robust validation and human oversight before unsupervised deployment.

Key Findings

  • GPT-4 diagnosed student errors in the Apprentice Tutor College Algebra ITS with an average accuracy of 87.8%, performing best on incomplete and correct answers and worst on "wrong field" errors and responses containing multiple errors, with accuracy heavily dependent on the problem context provided in the prompt.
  • Adding in-context information to the prompt — the tutor interface structure and Bayesian Knowledge Tracing skill estimates — dramatically improved error diagnosis, raising logical-error identification from 40% to 81% on factoring problems; the paper demonstrates that embedding LLMs within the structured framework of Intelligent Tutoring grounds generation and reduces hallucinated diagnoses.
  • Only 4.8% of generated feedback contained factually incorrect directions, and about 66% was targeted to the student's specific mistake, yet roughly 35% of hints were too general, incorrect, or bottom-out hints that gave away the correct answer.
  • Even when GPT-4 misdiagnosed an error, it still crafted relevant or general-but-correct feedback ~74% of the time by restating the concept or the expected answer format, suggesting the quality of corrective Feedback is partially decoupled from diagnostic accuracy.
  • LLM-based feedback evaluation (simulated-student adjustment and prediction tests) proved poorly aligned with human judgment: only 21.4% of hints passed both tests, targeted feedback passed just 30% of the time, and the metrics favored hints that simply gave away the answer — underscoring the difficulty of automated quality assurance for generated feedback.
  • Sentiment and readability analyses (TextBlob, CAREC, Coh-Metrix L2) showed most feedback was neutral-to-positive in tone and generally comprehensible, though a subset was dense or directive; the study did not measure actual learning outcomes with students.

Connected Concepts

Connected Articles

Citation

Reddig, J. M., Arora, A., & MacLellan, C. J. (2025). Generating In-Context, Personalized Feedback for Intelligent Tutors with Large Language Models. International Journal of Artificial Intelligence in Education, 35(4), 3459-3500.