3 days ago
AI teaching assistant shows real promise at S. Korea tech university KAIST
South Korea's top science and engineering university has rolled out a custom-built AI teaching assistant to help students in large graduate courses, and early results show it helped reduce repetitive student questions while encouraging more active, self-directed learning.
At Korea Advanced Institute of Science and Technology, a team of researchers led by AI graduate school professor Choi Yoon-jae and design professor Hong Hwa-jung developed a virtual teaching assistant, or VTA, that can answer student questions at any time, with responses tailored to specific lectures and coursework.
The tool was piloted last fall in a programming class for AI, taken by 477 graduate students. It's the first time such a system has been tested at scale in a Korean university setting.
What sets the KAIST VTA apart is that it isn't a generic chatbot. Instead, it runs on what's called a retrieval-augmented generation model, which pulls directly from course materials like slides, coding exercises and lecture videos. When students ask a question, the system finds the most relevant content and formulates a response based on that context. This means answers are grounded in what's actually taught, not just generated from a general AI model.
Over 50 percent of students used the system regularly during the 14-week semester, generating nearly 3,870 questions and answers. Students without a strong background in AI or coding were among the most active users, suggesting the VTA helped close knowledge gaps for those new to the subject. These figures come from internal usage data collected by KAIST during the semester.
The system didn't just benefit students. According to lead TA and doctoral researcher Kwon Soon-jun, it reduced the number of routine questions from students, such as basic concept definitions or explanations already given in class. That allowed human teaching assistants to focus more on deeper, more complex issues. Compared to the previous year's course, the volume of questions requiring direct responses from TAs dropped by around 40 percent, based on data compiled by Professor Choi.
Students also appeared more comfortable asking questions through the VTA than to human TAs, especially when it came to theoretical topics. Surveys conducted by Choi's research team before, during and after the course showed that students became increasingly confident in the system's reliability, and those who had been hesitant to speak up in class reported higher satisfaction levels when using the AI assistant.
The VTA's source code has been released publicly on GitHub to encourage adoption by other educators and researchers. The work was also accepted to the Industry Track at ACL 2025, one of the leading international conferences in natural language processing
