Adaptive Math Practice and Teacher Data Use: A Randomized Evaluation of AI-Enabled Learning in Uzbekistan

The intervention replaces one of five weekly Grade 5 mathematics lessons with a supervised Eduten session during regular school hours. Eduten provides curriculum-aligned adaptive exercises, immediate feedback, and structured practice. Treatment 1 receives the standard implementation model. Treatment 2 receives the same student-facing model plus structured support for teachers to review and use platform-generated information about student progress and learning gaps during their remaining mathematics lessons.

2026-09-02

2026-12-18

Student mathematics achievement at endline, measured by the overall IRT score on an independently administered Grade 5 mathematics assessment.

The independent assessment will cover foundational mathematics content from Grades 1–4 and Grade 5 curriculum content expected during the intervention. Item responses will be scored using an item-response-theory model. The overall endline IRT score will be oriented so that higher values indicate stronger achievement and standardized for interpretation in standard-deviation units using the control-group endline distribution. The corresponding baseline IRT score will be used as a precision-improving covariate.

1. Foundational numeracy IRT score (the items from the overall assessment that will contribute to this sub-score will be pre-registered ahead of the endline data collection)
2. Grade 5 curriculum-aligned mathematics IRT score (the items from the overall assessment that will contribute to this sub-score will be pre-registered ahead of the endline data collection)
3. Common Eduten quiz or assessment scores among the broader Grade 5 population in Treatment 1 and Treatment 2 schools.
4. Teacher accuracy in assessing student performance.
5. Observed instructional practices and data-informed instruction.
6. Teacher-reported data use, workload, stress, and professional judgment.
7. Student mathematics motivation, anxiety, frustration, discouragement, embarrassment, perceived labeling, and perceived teacher attention.
8. Implementation and fidelity measures from platform and field data.

The two independent-assessment domain scores will be constructed using prespecified item mappings and IRT scoring. The items from the overall assessment that will contribute to each of these two sub-scores will be pre-registered ahead of the endline data collection. Teacher prediction accuracy will compare class- or student-level predictions made before endline results are available with realized endline performance. Classroom-practice outcomes will use selected TEACH modules and clearly separated study-specific indicators. Related survey items may be combined into standardized indices, with item membership and sign conventions fixed before treatment-effect analysis. Platform outcomes will use a common assessment presented comparably in Treatment 1 and Treatment 2. Because control students will not use Eduten, these platform outcomes will be used for the Treatment 2 versus Treatment 1 contrast only.

The study is a three-arm, school-cluster randomized controlled trial in 450 eligible public schools in Tashkent City and Tashkent Region. Schools will be assigned in equal proportions to business-as-usual control, Eduten adaptive mathematics practice, or Eduten plus structured teacher use of platform-generated learning data. One Grade 5 class per school will be randomly selected for the independent baseline and endline assessments, teacher survey, and classroom-level measurement. The main analysis will estimate intention-to-treat effects based on school assignment.

Not available

Randomization will be performed by computer. Schools will be randomized within prespecified blocks defined by geography and school size and, if available and sufficiently complete, prior school performance. The randomization code, input frame, seed, block definitions, and assignment output will be archived. Within schools with multiple Grade 5 classes, survey software will randomly select one eligible class on site.

School. One Grade 5 class per school will subsequently be sampled for intensive independent measurement. Treatment is assigned to the school and applies to the Grade 5 population in treatment schools.

Yes

450 schools

Approximately 12,150 students in the independent assessment sample, assuming one randomly selected Grade 5 class averaging 27 students in each of 450 schools. Platform-based outcomes are expected to cover approximately 18,000-22,500 Grade 5 students across the 300 treatment schools, depending on realized Grade 5 enrollment and data availability. Approximately one Grade 5 mathematics teacher and one sampled class will contribute teacher- and classroom-level measures per school.

150 control schools, 150 Treatment 1 schools, and 150 Treatment 2 schools. The independent assessment is expected to include approximately 4,050 students per arm if sampled classes average 27 students. Platform outcomes will be available only in the two treatment arms and may cover approximately 60–75 Grade 5 students per treatment school.

Power calculations assume 150 schools per arm, 80 percent power, a two-sided 5 percent significance level, an intraclass correlation of 0.20, a coefficient of variation in cluster size of 0.50, and equal allocation across three arms. For the independently administered baseline and endline assessments, approximately 27 students will be assessed per school. The corresponding minimum detectable effect size is approximately 0.16 standard deviations without precision gains from baseline data, 0.15 standard deviations if baseline covariates explain 5 percent of outcome variation, and 0.15 standard deviations if they explain 15 percent. For the common platform-based assessments administered to approximately 75 students per school in the two treatment arms, the minimum detectable effect size is approximately 0.15 standard deviations without baseline precision gains, 0.15 standard deviations with 5 percent baseline explanatory power, and 0.14 standard deviations with 15 percent baseline explanatory power. Thus, the study is designed to detect pairwise effects of approximately 0.14–0.16 standard deviations, depending on the outcome sample and the predictive power of baseline data. These calculations apply to unadjusted pairwise comparisons and will be updated using realized cluster sizes, outcome availability, and baseline explanatory power.