AI-Powered Tutoring: Unleashing the Full Potential of Personalized Learning with Khanmigo

Educational disparities rooted in socioeconomic status (SES) continue to be a major public policy issue, with significant impacts on long-term economic opportunities for young people. Traditional classroom settings, characterized by a wide range of academic levels among students, present substantial challenges to effective teaching and learning. Teachers cannot easily navigate diverse individual educational needs while being required to present a fixed set of topics over a limited period of time. Variance in academic preparedness and learning comprehension means that advancing new material can leave some students struggling to keep up, not due to a lack of capability but because of increasing foundational gaps. This heterogeneity is further exacerbated by factors such as the COVID-19 pandemic, which has widened the educational gaps and underscored the urgency of finding scalable solutions to provide personalized instruction.

Mastery learning, a pedagogical approach where students must demonstrate a high level of understanding before moving to the next topic, offers a framework for addressing these disparities. However, the traditional model of one-on-one tutoring, while effective, may be too costly and operationally challenging to scale. This project aims to explore the roles that Computer Assisted Learning (CAL) and Large Language Model (LLM) technologies might play as scalable solutions to these challenges to support mastery learning across diverse educational settings.

The main intervention is a program to support regular CAL practice for 30-45 minutes a day during a school's "What I Need" time at the end of each school day for Tier 2 and 3 Grade 6-6 students. Randomization is at the grade level across approximatley 50 schools. Schools receive training, support, and weekly guidance around successful implementation, with a focus on ensuring high dosage practice and adequate progress. Students follow an incremental roadmap of mastery learning both for math and English, with teachers and supervisors choosing particular units and topics for students to follow. Coaches help teachers and administrative staff ensure fidelity.

2024-08-01

2025-05-31

Primary outcomes are math and English assessment scores. Standardized tests are collected multiple times during the school year and at the end. These outcomes will be normalized so that the control group variables have mean zero, standard deviation one for each grade. We aim to also collect subjective survey data from both control and treated students around math and english satisfaction, attitudes towards school and life in general.

Subjective questions will be measured on a 7 point Likert scale.

We will also examine intermediate outcomes, looking at impacts on Khan Academy practice time and progress. These outcomes will be measured in terms of total practice minutes and proficiency levels attained over the school period. In follow up years we will also examine second year impacts on similar assessment scores to first year primary outcomes.

Grade 6-8 students classified as Tier 3 at the beginning of school similar Tier 2 students that schools that choose to include in the program would be the target population. Grades/schools would be randomized to determine which grades will receive the program and support, and which ones will be selected as comparison grades. Each school would have at least one grade selected, and not more than two. Grades selected for randomization will have their dedicated daily "What-I-Need" time used for Khan Academy practice both for math and ELA. Teachers would receive mandatory training before the beginning of school with simplified instructions and a project manager would be hired to manage khoach meetings and field visits to help ensure fidelity. Supervisors during WIN time will meet regularly (weekly) with khoaches or RTI leads to assess progress. All students would be given khanmigo access to maximize power for estimating the full program. The analysis will include impacts both at the end of the program in year 1, and a year after the program has concluded, at the end of year 2.

Not available

Randomization using R code with seed is:

1) choose random number between 0-1 for each school. Above or below 0.5 determines if school has 1 or 2 treated grades.
2) choose random number between 0-1 for each grade in each school, sort, and top 1 or 2 grades, depending on (1) will be treated grades.

Grades are randomized within each of 22 schools. Each school will have either one or two grades treated (out of 3, for Grades 6-8)

Yes

22*3 = 66 for the first year. Second year analysis will include some students treated once, some twice, and some not at all, which will help increase power and examine for multiple dosage analysis.

66 grades, with about 50 Tier 2 and 3 students in each grade (at total of 3,300 students in Year 1).

33 treated grades

With ICC of 0.1 (in part from conditioning on previous test scores), MDE is 0.17 for 80% power at 5% statistical significance.