LLM use during exam preparation and student performance in university mathematics: a quantitative observational study

Resumo

The integration of large language models (LLMs) into student study practices raises important questions about their association with academic performance, yet empirical evidence from authentic university settings remains scarce. This observational study examined whether self-reported LLM use during exam preparation was associated with performance in a first-semester university mathematics course focused on discrete mathematics and mathematical logic. Data were collected from N=197 students who completed two midterm tests and indicated after each test whether they had used an LLM during preparation. A linear mixed-effects model with a random intercept for student was applied to account for the repeated measurement structure and unbalanced participation in the second test. Reported LLM use was associated with an average score increase of approximately 11 points (95% CI [3.82, 19.12], p=.003), with no significant interaction between LLM use and test (p=.907), indicating a consistent association across both assessments. Participation in the second test was strongly predicted by first-test performance (d=0.97), while LLM use was not a significant independent predictor after controlling for prior achievement. These findings contribute quantitative evidence to ongoing discussions about the role of generative AI in university mathematics education.

Dados da obra

Tipo

Artigo

Periódico

Frontiers in Education, v. 11

Editora

Frontiers Media S.A.

Ano

2026 (2026-07)

Idioma

en

Palavras-chave

academic performance; exam preparation; generative AI; large language models; mathematics education; mixed-effects model

Temas

Avaliação

DOI

10.3389/feduc.2026.1875391

ISSN

2504-284X

Acesso

Acesso aberto

Fontes

DOAJ — Directory of Open Access Journals

Na Biblioteca desde

01/10/2026 · verificada em 03/10/2026