Training a QUAN Model as a tutor

Laurent Béland, Smith Engineering This is a ChatGPT4-based chatbot being used in MECH 203. This class has roughly 280 students.

Article Category

AI-Augmented Learning Activities

AI integration

Assignment

ChatGPT

We trained a QUAN model (ChatGPT 4-based chatbot, retrained on course content and provided prompts to act as a tutor) as a tutor.

The idea was that the AI tutor could act as a “fancy solutions manual”, providing instant feedback to students, and get them “unstuck” quickly.

There are benefits to combining in-person problem solving with AI assistance.

Even later in the semester, when use of the AI tutor is not required, many students are still using QUAN to help them understand how to solve the problem sets.

Overall, I think this was successful. The students were asked to use this AI tutor to help them solve the problems in the problem set. We performed a two-section study (~140 students in each section). One section (control) had regular lectures in-class, and problem solving at home. The other section had online lectures at home and problem solving in-class (intervention). In week 1 (pen-and-paper math) control got 60% in the quiz and intervention got 68%. In week 2 (numerical methods), we flipped the control and intervention section. Control got 85% on the quiz and intervention got 90%. To be clear, both sections had access to the AI tutor when solving their problem set. In the future, I think that I’ll stick to flipped classroom.

Additional Information

The students were asked to use this AI tutor to help them solve the problems in the problem set.

Flipped classrooms.

This activity could be used in a wide range of courses where students benefit from structured practice, immediate feedback, and opportunities to work through problems actively. In our case, it was implemented in a large introductory undergraduate course on mathematical, computational, and statistical tools for engineers, with approximately 300 students divided into two sections of about 150 students each. The main takeaway was that a flipped-classroom format, supported by an AI tutor, could outperform a more passive lecture-based format even at this large scale. While our implementation was in a relatively introductory quantitative course, my understanding of the broader teaching literature is that flipped and active-learning approaches can be effective in both foundational and more advanced courses, including smaller classes.

The learning objectives addressed include developing problem-solving fluency, applying mathematical and numerical methods to engineering problems, identifying and correcting errors in solution strategies, and learning how to use AI tools productively as a source of feedback rather than simply as a source of answers. The main pedagogical goal was to help students get unstuck while working through problems, while using in-person class time for active problem solving rather than passive content delivery.