Meet Arthur: Your AI Virtual Teaching Assistant

How often have you experienced the frustration of waiting for a response from your professor about a homework question? Or, if you are an instructor, how frequently do you find yourself overwhelmed by the sheer volume of student inquiries on the same homework problems?

At large, public institutions like Purdue University, a single professor typically oversees a substantial number of students with only a limited number of teaching assistants for support. When a student encounters difficulty with an assignment or errors in their homework, they must await guidance from an instructor. This usually occurs during office hours on a different day or via email after several hours, disrupting the continuity of the student's learning process. As the homework deadline approaches, the volume of student emails swells, packed with similar questions. This not only burdens the instructor's already heavy workload but also delays the turnaround time for providing much-needed feedback. Consequently, instructors often find themselves repeating the same answers multiple times. This situation is far from ideal for both parties involved. Students lose precious time that could be spent consolidating their understanding, and instructors are unable to efficiently manage their time and resources.

Since last year, students and instructors in IE343, an Engineering Economics course, have been introduced to a novel alternative: Arthur – an AI-based virtual TA[1] developed by our research project[2]. Arthur was named in honor of Arthur M. Wellington, a trailblazing American civil engineer who made pioneering efforts in the field of engineering economics. Interacting with AI Arthur is not only straightforward but also effective. When students submit their final answers to a practice problem, Arthur immediately analyzes them, pinpointing likely errors and offering instantaneous feedback. With swift feedback, students can correct their problem-solving approach in real time and maintain their learning flow without the pain of lengthy waits.

To develop our AI teaching assistant model, we utilized historical graded submissions, including homework, quizzes, and exams, from previous semesters. Unlike the traditional tutoring systems that heavily rely on instructors’ expertise to identify mistakes, we found that the historical graded submissions themselves tell more stories. They narrate common patterns and unique errors among student cohorts, offering deeper insights than previously recognized. The data fueling our AI model’s training was accumulated passively in previous semesters. With the widespread use of digital learning platforms like Gradescope, Brightspace, and Variate at Purdue, a continuous stream of valuable data is generated every day in every course. This wealth of information forms a rich data mine, facilitating the development of an effective AI teaching assistant for courses, especially in STEM, avoiding the high labor costs typically associated with purpose-built training data creation.

Our primary goal with the AI model is to replace the labor-intensive process that requires instructors to investigate, identify, and communicate student errors with a more efficient process that can predict the most likely mistakes and guide the student toward immediate, effective revisions in their approach. Since its creation, Arthur has been implemented for three consecutive semesters in IE343. A survey conducted in the Fall 2022 IE343 class revealed Arthur's positive impact on student performance. Regarding the perception of Arthur’s helpfulness, there was a statistically significant consensus among students who used this tool: they agreed that Arthur was instrumental in enhancing their confidence in the course. It helped maintain their learning momentum and mitigated the delay in receiving instructional feedback. Furthermore, in terms of actual grade improvement, even though using Arthur was optional, students who engaged with it during their studies achieved significantly higher scores compared to their peers who opted not to use the tool.

Arthur was created before the advent of the large language model. Looking forward, this opens up exciting possibilities to leverage the generative capabilities of these advanced models for the AI TA tool. Envisioning a future where learning is not only more personalized but also readily accessible to students from diverse backgrounds and with varying learning needs, we are enthusiastic about the potential of this AI-powered TA implemented in more courses. Such a virtual assistant promises to revolutionize the educational landscape, making learning more adaptable and inclusive in higher education.

[1] AI is short for Artificial Intelligence and TA represents Teaching Assistant.

[2] A teaching innovation project funded by Purdue’s Innovation Hub

About the Author: 

Zhuoli Yin is a fourth-year PhD student in the Edwardson School of Industrial Engineering at Purdue, working with Prof. Hua Cai. His research interest lies in the analysis, computation, and decision-making in urban computing to pave a sustainable path for our cities, by leveraging techniques from artificial intelligence, operations research and life cycle assessment. His research has diverse applications, including shared urban mobility, net-zero carbon emissions and intelligent tutoring systems.