← গবেষণা সংবাদে ফিরুন
A diverse mathematics class works with an AI tutoring system while a teacher compares student reasoning and assessment evidence
রিভিউEvidence synthesis2026৯ জুল, ২০২৬· 3 min

Review of 12 studies finds promising mathematics-tutoring gains while generative-AI achievement evidence remains thin

Neo Molemane, Moeketsi Mosia, Felix O. Egara

Discover Education

৫০০-শব্দের সারাংশ

A diverse mathematics class works with an AI tutoring system while a teacher compares student reasoning and assessment evidence

Molemane, Mosia and Egara review empirical evidence on AI-based tutoring and assessment systems in mathematics education. Following PRISMA procedures, they searched Scopus and Web of Science for studies published from 2015 through 2025. The search produced 1,749 records, 76 full texts were assessed, and 12 studies met the final inclusion criteria. Two reviewers independently screened, extracted and appraised the evidence, resolving disagreements through discussion. Methodological quality was examined with the Mixed Methods Appraisal Tool, while findings were synthesized narratively because interventions, learners and outcomes were too heterogeneous for a defensible pooled effect estimate.

Across the included studies, AI tutoring and assessment systems generally showed promising associations with mathematics achievement, engagement or personalized support. Several reports suggested that lower-performing learners could benefit particularly from adaptive guidance, immediate feedback and practice matched to current understanding. The stronger claims came from experimental or quasi-experimental work rather than perception surveys. Effects were not uniform: they varied with learner characteristics, instructional context and implementation fidelity. A system's label as artificial intelligence therefore did not determine its educational value; the design of feedback, teacher integration and opportunities for students to reason remained important.

The review draws a useful distinction between established tutoring systems and recent generative AI. Studies involving ChatGPT or comparable tools mainly reported positive perceptions, usability or engagement. Direct evidence that these tools improved mathematics achievement was still limited. The review consequently calls for larger, more rigorous trials and clearer accounts of intervention mechanisms. This boundary matters because a fluent explanation, a learner's positive rating and a correctly completed exercise are different outcomes. None alone demonstrates durable conceptual understanding, independent problem solving or transfer to an unfamiliar mathematical task.

The evidence base is small and diverse. Twelve studies cannot support a universal estimate across ages, curricula, countries and AI architectures, and the narrative synthesis provides no overall percentage gain. Restricting the search to two databases may have missed eligible work, while publication and language choices can shape what was found. Some included studies were authored by members of the review team, a relationship the paper acknowledges. The journal also identifies the available article as an early-access manuscript that had not completed final editorial processing, so minor presentation or metadata changes could follow.

For Hong Kong, the review supports carefully designed trials rather than immediate system-wide procurement. A school could compare an AI tutor with existing teacher-led practice, map every task to the local curriculum, and examine outcomes separately for learners beginning at different achievement levels. Assessment should include an unaided near-transfer problem, delayed retention and an explanation scored without knowing the student's condition. Chinese and English mathematical language, accessibility, device access, teacher workload and the consistency of feedback across common misconceptions should be recorded alongside scores.

The paper's defensible conclusion is that AI tutoring and assessment in mathematics is promising but conditional, while the achievement evidence for generative AI remains notably thin. Decision-makers should ask which mechanism produced a result, for whom, under what instructional conditions and for how long. Publishing implementation fidelity and null findings would make local evidence more useful than a feature demonstration. Until larger comparative studies accumulate, the review is best used to design stronger evaluations, not to certify every AI mathematics product as effective.

সম্পর্কিত পেপার

Illustrative photograph of two educators examining a simple prediction baseline and uneven error bars on a seminar-room display.
টুল / ডেটাসেট2026
টুল / ডেটাসেট 245

Preprint: Tutoring logs add little to simple score predictions in one course

Zeyu Dong, Mikael Kwame Kuada, Asa Holley Buchakjian-Tweedy, Ibrahem Aljabea

arXiv preprint

A retrospective analysis compares tutoring-interaction features with constant score predictions. Small overall gains and uneven errors show why a support model needs more than a favorable average accuracy figure.

preprintlearning analyticsprediction baselines
৫০০-শব্দের সারাংশ পড়ুন →
Illustrative research-review scene with two raters comparing repeated and progressive activity sequences beside a screen with one rising and one falling chart.
টুল / ডেটাসেট2026
টুল / ডেটাসেট 239

Preprint: Raising a tutor metric lowered human ratings in an optimization experiment

Daniel Domínguez Figaredo, Rafael Fernández De la Cruz

arXiv preprint

An optimization experiment produced higher automated scores but lower human ratings for instructional sequences. Its synthetic cases and constrained intervention make it a warning about evaluation design, not evidence of harm to actual students.

preprintAI tutoringevaluation validity
৫০০-শব্দের সারাংশ পড়ুন →
Illustrative scene of an educator and learner discussing a four-part tutoring loop and a balance-scale drawing on a glass board in a bright study area.
রিভিউ2026
রিভিউ 237

Preprint: A tutoring-loop review separates convincing dialogue from evidence of learning

Sayantan Adak, Rima Hazra, Koyel Mukherjee, Digbalay Bose, Somnath Banerjee

arXiv preprint, version 2

A review examines how learner information, instructional choices and evaluation connect in AI tutoring. It highlights the need to inspect what students subsequently do, beyond the apparent quality of a conversation.

preprintAI tutoringlearner modeling
৫০০-শব্দের সারাংশ পড়ুন →