Original title: Large Language Models for Test-Enhanced Learning: A Tutorial and Proof of Concept
Ariana Modirrousta-Galian, Daryl Yu Heng Lee, Frederic Dick, Eva Lash +1 · September 2026
Testing yourself helps you remember what you study, a benefit called test-enhanced learning. However, making practice questions takes a lot of time. Researchers tested whether artificial intelligence tools like ChatGPT and Gemini could make these tests and help students learn. In the experiments, simple multiple-choice quizzes with standard feedback did not help people remember a video two days later. But when the computer models created short-answer questions and gave personalized feedbackFeedback that responds to the specific answer a student gave, not a one-size-fits-all message. on the answers, students remembered the material much better after seven days compared to just rewatching the video. This worked even if the final test used a different question format. In a real college course, taking these AI short-answer tests every two weeks was linked to higher grades, and students viewed the tool favorably.
Imagine studying for a geography quiz by reading the atlas over and over. The pages start to feel familiar, but a week later the names slip away. Now suppose a friend instead hands you a blank map, asks you to fill in the countries, and then tells you exactly which ones you mixed up and why. In this paper, the researchers tested whether AI-written practice questions work like that blank-map exercise: questions that need a written answer and come with an explanation tailored to your mistakes helped people remember a video a week later, while re-reading the material or using questions with a simple right-or-wrong answer did not.
Students and teachers both know practice tests help, but writing good questions and feedback takes time. This paper shows that large language models can do that work, making regular retrieval practice feasible in real classrooms. The catch: only short-answer questions with personalized feedback helped; multiple-choice with fixed feedback did not.
Because the classroom study was correlational, it cannot show that the AI-written tests caused better grades. The experiments used only psychology undergraduates, two LLMs, and recall measured after two or seven days, so longer-term or other-population effects are unknown. The null result for multiple-choice also means the benefit is not guaranteed across all question formats.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Large Language Models for Test-Enhanced Learning: A Tutorial and Proof of Concept — Ariana Modirrousta-Galian, Daryl Yu Heng Lee, Frederic Dick et al. — PsyArXiv — doi:10.31234/osf.io/h5r4w_v1
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.