As the country continues to grapple with post-pandemic dips in children’s math and reading scores, University of Utah education professor Chenglu Li and his collaborators are combating that trend with an AI-powered app that has already impacted thousands of students. The program, ALTER-Math, which stands for AI-augmented Learning by Teaching to Enhance and Renovate Math Learning, turns middle schoolers into teachers: students earn points as they help fictional peers solve real-world algebra problems and understand concepts.

The project, which Li started working on just before finishing his educational technology PhD at the University of Florida (UF), was partially inspired by his experience as a peer tutor in high school math.

“I realized how much I learned by teaching my friends,” said Li, who joined the U’s Department of Educational Psychology as an assistant professor in 2023 and was named a One-U Responsible AI Initiative faculty fellow in late 2024.

Learning-by-teaching is an established framework in the learning sciences, shown to be effective primarily through studies in physical classrooms with human students. Accordingly, ALTER-Math runs on intentionally “weak” large language models, a type of artificial intelligence where algorithms use text data to understand human language and generate appropriate responses.

“The models are not weak in performance, but designed to foster learning by prompting students to think critically,” Li said. “The punchline here is: While AI has yet to become a reliable teacher, it is already an excellent student.”

ALTER-Math shifts students’ roles from passive learners to proactive teachers. Students guide an AI teachable agent—their fictional peer—to solve math problems in themes such as music and the environment, actively participating in math learning by prompting the agent, correcting its mistakes, and observing its growth.

Since its October 2023 launch, ALTER-Math has been used by more than 50,000 middle school students, more than half of whom come from low-income backgrounds. Researchers have quizzed 6,000 of those students before and after they used ALTER-Math or the business-as-usual online learning platform Math Nation, which hosts ALTER-Math. Students who used ALTER-Math scored higher than those who didn’t use it.

Specifically, accessing ALTER-Math tends to increase students’ learning gains, defined as post-quiz score minus pre-quiz score, by 1.56 times, Li said. “It’s encouraging to see how ALTER-Math makes learning fun and interactive—while also delivering real, meaningful learning benefits,” Li added. He and his collaborators will continue to refine the app until fall 2027, when the project concludes. The team hopes to show that ALTER-Math can double the rate of math learning.

This work, Li said, is highly relevant to the nationwide challenge to boost math and reading scores. According to the Harvard University-based Education Recovery Scorecard, as of spring 2024, U.S. students were “nearly half a grade level behind pre-pandemic achievement in math and reading.” And the achievement gap has widened: the difference in math scores between students in affluent and low-income districts grew by 11%.

Advanced learning tools such as ALTER-Math can help students across socioeconomic groups by “amplifying their agency and ownership” of the learning process, Li said. “We especially targeted middle school math as it’s the gateway to more advanced mathematical tools and STEM fields, which can also impact students’ future careers—and ultimately the nation’s workforce,” he added.

ALTER-Math is a large-scale collaboration involving six institutions: UF, the U, Vanderbilt University, Stanford University, Duke University, and Accelerate Learning, the parent company of Math Nation. Li’s role is akin to that of a chief technology officer and vice president of research. He serves as the managing director, co-leading the conceptualization, development, and research efforts alongside project lead Wanli Xing, Li’s PhD advisor from UF.

The team is supported by $10 million in funding from leading education philanthropists and organizations through the Learning Engineering Virtual Institute. And ALTER-Math has acquired a huge audience in just over a year thanks to integration with Florida public schools via Math Nation and the UF Lastinger Center for Learning.

In addition to directly improving student outcomes, ALTER-Math has contributed to the broader fields of learning analytics, AI in education, and large language model applications. The team has pre-trained and fine-tuned six large/vision language models, created three robust datasets with tens of thousands of entries, and published two open-source codebases.

As Li and his collaborators work toward their goal of reaching 200,000 students by 2027, they’re expanding into new states. They’re currently finalizing an agreement with the Alabama State Department of Education, and Li is actively working to introduce ALTER-Math to Utah schools, though it may be a long journey to district-level integration.

As one of the seven inaugural One-U Responsible AI Initiative fellows, Li is also leveraging the expertise of his cohort. He shared relevant ALTER-Math data with Ana Marasović to explore more effective problem generation using generative AI. He also hopes to leverage Vineet Pandey’s expertise in human-computer interaction to design responsible and effective AI solutions for K-12 education. “The fellowship not only provides invaluable support for recruiting talented doctoral students to advance my research agenda,” Li said, “but also offers access to a vibrant community of scholars with diverse expertise for meaningful collaboration.”

• Kelly Hermans Communications manager Scientific Computing and Imaging (SCI) Institute
  ‭(706) 296-8037‬ kelly.hermans@utah.edu