StemPal Brings Guided AI Support to STEM Learning

(Bethlehem, PA) Lehigh University College of Education researcher Zilong Pan is developing an AI-powered AI agent suite to help students work through STEM learning tasks and problems while keeping teachers in the learning loop.

When a student gets stuck on a math problem, computer program, or statistical analysis, generative artificial intelligence can produce an answer almost instantly. Pan is interested in a different question: What if AI helped students figure out the answer themselves?

Pan is developing StemPal, a suite of educational AI agents that provide students with guided hints and instruction while giving teachers data and analytical tools to better understand student progress and engagement.

Rather than serving as a shortcut around learning, StemPal is designed to make AI a collaborative partner, or pal, during the learning process.

“I don’t want AI to simply give students an answer,” Pan said. “The goal is to provide the right amount of support at the right time so students can continue thinking, questioning, and solving the problem themselves.”

That philosophy reflects Pan’s broader research at Lehigh. A faculty member in the College of Education’s Teaching, Learning and Technology program, Pan studies emerging educational technologies, personalized learning environments enhanced by artificial intelligence, immersive learning experiences, and instructional decision-making informed by learning analytics.

StemPal currently brings together AI-supported tools for several STEM disciplines.

MathPal provides guided problem-solving hints, step-by-step instruction, and opportunities for students to learn mathematical concepts through guided practice. StatPal supports statistical learning through contextual hints and guidance during data analysis, while HackPal provides programming hints and debugging assistance intended to help students understand coding concepts rather than simply repairing their code for them.

The distinction is important in an era when students have easy access to general-purpose generative AI tools capable of producing completed solutions.

With StemPal, the emphasis is on scaffolding: providing students with enough support to move forward without removing the intellectual work that leads to learning.

“Students are going to use AI,” Pan said. “Education has to move beyond asking whether they should use it and begin thinking carefully about how we design AI experiences that help students learn. There is a major difference between an AI system that completes a problem for you and one that helps you develop the skills to complete it yourself.”

The tools are also designed with teachers in mind. StemPal collects information about student interactions and progress and presents the processed information on a teacher-dashboard, creating opportunities for educators to see where students struggle, what kinds of assistance they seek, and how they approach problems.

That teacher-facing component reflects another central principle behind Pan’s research: AI should augment teachers rather than limit their role of instruction.

MathPal, the most extensively studied component of StemPal, has already moved beyond the development stage and into classrooms.

A study led by Pan and collaborators examined how MathPal, which was designed to provide high school students with both conceptual support and metacognitive guidance, helped them think not only about a mathematical problem but also about their own problem-solving process.

MathPal operates as a browser extension compatible with digital math platforms. Students can ask questions and receive explanations, practice problems, strategies, and step-by-step guidance without the solution being revealed immediately. Its design is informed by principles of metacognition and a growth mindset, encouraging students to persist when they encounter difficulty.

Researchers conducted two rounds of usability testing in 2024. The first involved a ninth-grade math teacher, a school district professional development and technology integration supervisor. The second round involved three ninth-grade math teachers and 78 high school students who used MathPal in their classrooms for a month.

Teachers in the first round reported that MathPal provided clear mathematical explanations, generated useful performance information, and offered reasonable problem-solving strategies. Their feedback also led the development team to strengthen user management, data security, and the system’s ability to redirect students when conversations moved away from mathematics.

Student feedback provided another perspective.

Students reported that MathPal was especially useful when working through difficult problems because it broke the process into manageable steps. They also said it could provide support when a teacher was helping another student, introducing them to different ways of organizing and approaching problems. Researchers found that students also became better at communicating with AI, learning to formulate more precise questions to receive more useful responses.

For Pan, that last finding points toward another emerging educational need: AI literacy.

“Knowing how to ask good questions, evaluate the response, and determine what you still need to understand is becoming an important learning skill,” Pan said. “We want students to be active participants in that process, not passive recipients of whatever an AI system produces.”

Pan and his collaborators are also studying what happens within conversations between students and MathPal.

In another study, researchers examined 48 ninth-grade Algebra I students who used MathPal over a 14-week semester. The students generated 1,214 conversational threads with the system, providing researchers with a detailed picture of how students seek assistance from an AI learning tool.

The study found that students most often approached MathPal for help with solutions or computations. MathPal, in turn, most often responded with strategic scaffolding, guiding students through the steps of solving a problem. Researchers also identified four distinct patterns of student-AI interaction, suggesting that students do not all use an AI tutor in the same way.

That finding could have implications for the next generation of educational AI.

A student who repeatedly asks how to take the next step may need a different type of support than a student who asks conceptual questions about why a mathematical principle works. Researchers concluded that AI learning systems could become more effective by recognizing those differences and adapting the type of scaffolding they provide.

The same interaction data could also help teachers identify students who are struggling, who rely too heavily on procedural assistance, or who fail to engage deeply with a concept.

That is where Pan sees the relationship between artificial intelligence and learning analytics becoming especially valuable.

“The technology can support the student in the moment, but it can also help the teacher understand what is happening across the classroom,” Pan said. “A teacher might not be able to observe every decision every student makes while solving a problem. These interactions create data that can help teachers identify patterns and make better-informed instructional decisions.”

Research on MathPal has also reinforced the importance of teacher oversight.

During usability testing, teachers recommended features that would allow them to control when students could access MathPal, upload their own instructional materials, and more easily categorize and analyze student interactions. The researchers concluded that teacher-in-the-loop mechanisms are essential for AI tools to remain aligned with classroom goals while allowing educators to intervene when necessary.

That approach differentiates StemPal from the idea of AI as an autonomous tutor replacing human instruction.

Instead, Pan envisions a partnership in which technology provides students with immediate support and teachers retain control over instruction.

The work also aligns closely with Pan’s broader objective as a researcher: designing technology-rich learning environments that respond to differences among learners rather than expecting every student to learn in the same manner.

As StemPal continues to develop, research on MathPal is providing Pan and his collaborators with something particularly valuable for educational technology: evidence from students and teachers using the tools in real classrooms.

“We’re not developing the technology first and then asking educators to figure out what to do with it,” Pan said. “Teachers and students are helping us understand what works, what doesn’t, and what needs to change. That has to be part of the design process.”

As generative AI becomes an increasingly common presence in education, Pan believes that distinction will matter.

“The question isn’t simply whether AI belongs in education,” Pan said. “The more important question is what kind of learning we want AI to support. With StemPal, we’re trying to ensure the technology helps students become better thinkers and problem solvers while giving teachers another tool to understand and support their learning.”

About Lehigh University College of Education

Lehigh's College of Education offers top graduate programs that focus on impactful research, transformative learning, and community partnerships, both locally and globally.