A pilot AI tool generates math clues to help students solve problems

As a middle school math teacher, Alfons Prince is used to having a lot on his plate.

“You’re one teacher in a class of 25 kids,” says Prince, who teaches seventh and eighth grades at Center City Public Charter School in Washington. “When a student gets stuck on a problem, it’s hard for me to stop everything and give them one-on-one attention.”

Balancing the individual needs of students with the demands of the entire class is a perennial dilemma for teachers. But computer science researchers from Virtual Institute for Learning Engineering we believe that artificial intelligence technology can be part of the solution. “If ChatGPT and tools like it can provide students with more immediate, personalized help, teachers will be able to focus where they are needed most,” says Zachary Levonian, senior machine learning engineer at LEVI.

Tutoring is one of the most effective forms of support for students, but high staff costs make it difficult to scale. Virtual math tutors offer an alternative, but existing systems tend to be inflexible and narrowly focused. If a student misunderstands a problem, the system offers a pre-set sequence of prompts to help get them back on track.

The problem, says Levonian, is that children can stumble in many different ways, and the system only has one way to solve the problem.

Levonian and the LEVI team wanted to build a more versatile tool. One that could tailor its response to the specific type of error the student was making. They focused on large language models (LLMs), which excel at interpreting general prompts and generating appropriate responses. “We thought LLM programs like ChatGPT could be a one-size-fits-all solution,” says Levonian. “They give teachers the ability to tailor high-quality feedback to a wide range of student issues and responses.”

In June, they released a ChatGPT-based prototype called the Math Hint Generator Chatbot. The demo tool pulls material from verified online sources and a digital textbook. It is able to record sequences of targeted tips for many common errors.

The initiative is supported by the Walton Family Foundation, which is part of our ongoing efforts to seek new ideas in education and support innovators with creative solutions.

Prince was one of the tool’s beta testers. He considers himself quite tech-savvy. Prince, however, was cautious about introducing technology into his classroom. “I want to make sure my students know how to do math,” he says. “I’m afraid some of it may be lost if they take a technological shortcut.”

Still, Prince was intrigued by the prospect of having a tool that could “help kids get out of trouble without having to wait for a teacher.” Once he gained access to the demo, he entered real math problems and incorrect answers from his students to see how the chatbot performed. “It was very good in technical mathematics,” says Prince. “I created a problem of finding the volume of a cone. He used the formula correctly and in a way that could help build understanding of it for the student using it.”

While Prince was impressed with the quality of most of the prompts, he says the tool still requires a teacher’s watchful eye to check the results. “I could definitely see how it could be used to prepare for potential questions from kids,” Prince says. “Even though I have a developed curriculum, children are unpredictable. It would have saved me a lot of time if the bot had generated hints in advance about expected questions. I know a lot of teachers could benefit from this.”

While an early demo is underway, Levonian says the pilot is intended to give people the opportunity to brainstorm ways these tools can make their jobs more efficient. Meanwhile, he agrees that the best use of the current tool is to help teachers prepare lessons. Other practical applications could include using the chatbot to create a sequence of directions for online classes, or simply sparking ideas for how teachers can more actively teach children how to avoid common mistakes.

Looking to the future, the LEVI Lab AI team plans to integrate an updated version of the tool with Rori, the virtual math tutor they use in Africa. Ultimately, they hope to make it available directly to children so they can get help at any time, at school or at home.

“Before we can do this, we need to take a step forward to create language that is classroom-appropriate and culturally sensitive,” says Levonian. “They would need to know how to respond effectively to situations that teachers encounter on a daily basis.”

Future generations of the Math Prompt Generator Chatbot can be adapted to different school systems, curricula, and student populations.

“I’m very optimistic about this job,” says Levonian. “We hope this will help teachers understand the possibilities of using tools like ChatGPT. Nowadays, teachers spend a lot of time writing lessons. This means they don’t have time for other important things. We are really excited about the prospect of generating prompts that will save teachers time while making lessons more personal and responsive for children.”