�鶹��Ƶ

The �鶹��Ƶ is one of seven teams chosen to join the new (LEVI) from over 100 competing teams. The institute's goal is to double the rate of middle school math learning within five years, focusing on students from low-income backgrounds. 

Rates of mathematical achievement are now the lowest they have been in decades. Between 2019 and 2022, National Assessment of Educational Progress mathematics assessment scores declined 8 percent for all eighth graders, with only 26 percent of students scoring as proficient. 

Professors Sidney D'Mello and Tamara Sumner of the Department of Computer Science and Institute of Cognitive Science join professors Peter Foltz, Jennifer Jacobs and Jeffrey Bush of the Institute of Cognitive Science in leading the project team. 

 "The LEVI initiative is incredible,” D'Mello said. “It enables our 20-year dream to scale our work to impact hundreds of thousands of students." 

The project also involves a partnership with , a large non-profit provider of tutoring services to low-income youth and a collaboration with Professor Jacob Whitehill from Worcester Polytechnic Institute.

Transforming tutoring

CU Boulder's project is the Hybrid Human-AI Tutoring (HAT) platform. It aims to transform tutoring from only a human or technological solution into human and computer synergy that will reach over 275,000 diverse, low-income students over the next five years.

"We're going to blend what computers do best – finding patterns in large volumes of data – with what humans do best – nurturing and caring for each other," D'Mello said. 

D'Mello, who also directs the $20 million federally funded National AI Institute for Student-AI Teaming (iSAT), has considerable experience with connecting AI to student success. Several of the key technologies from iSAT and related projects will be leveraged to jump start the LEVI Institute.

The program also builds on the success of Saga, a highly successful "high-dosage" human tutoring program that lets students experience the equivalent of a second math class every day. 

While several external studies have shown that Saga doubled math achievement for low-income students. This initiative will address some of the key challenges that face the program. 

"We’re addressing the central challenge of scaling high-dosage tutoring while maintaining quality of service," D'Mello said. 

A problem of scale

Tutors are, in general, chosen for their subject-matter expertise, not their teaching experience or experience with low-income, school-age children. 

This lack of experience on the part of the tutor can lead to unequal outcomes for students. It also requires intensive and expensive on-the-job learning and targeted feedback to improve tutor performance and student engagement, something that is challenging to provide consistently. The HAT platform helps with this. 

HAT involves advanced computational machinery, including finely-tuned speech recognizers and large language models, combined with powerful analytical tools and reinforcement learning. 

Historically, artificial intelligence and machine learning have been plagued with bias. HAT, however, is deeply focused on removing bias from machine learning models and improving the emotional and social environment that tutors provide to their diverse cohort of learners.

The large language models will be enhanced to support multiple languages, speech vernaculars and styles, and nonverbal cues. 

HAT also carefully attunes to tutoring conversations and nonverbal signals, providing feedback during and after each session to ensure equitable learning opportunities for students across economic and racial demographics.  

A human-centered approach

As well as improving the quality of tutoring, the platform works to support the process of engaging students. Before the tutoring session, HAT helps the tutor determine what tasks would be helpful for the students in their groups. 

"HAT uses students' behaviors in past sessions to help tutors select challenging math tasks tailored to enhance learning for each small student group," D'Mello said. 

The advanced computation behind HAT also makes it possible for the platform to give tutors specific feedback on how to to engage students in collaborative, meaningful mathematical discussions. 

For example, HAT might prompt a tutor to ask students to provide reasoning for an answer they just gave, in real-time, to support their learning. 

HAT also aims to create personalized computational models. This means the model can, among other scenarios, ensure that behaviors from disabled and neurodiverse learners and tutors are accommodated for and not assumed to be non-productive. 

The final part of HAT is between-session coaching, where data about the tutoring session allows tutors and their human coaches to create and track goals towards highly effective tutoring . Over time, this is expected to improve tutor performance, student engagement and educational outcomes for students. 

Within five years, the solutions generated by LEVI teams are projected to reach millions of students.