In recent years, the computer science and engineering offerings at Avon Old Farms have been expanded substantially. Despite this, some of our students, like Junbum Cho ’26, still manage to go above and beyond what is required of them in the classroom. In his latest personal project, Junbum published a paper exploring some novel functions possible in the future of artificial intelligence (AI).

It began as a submission for a coding competition in his native South Korea. Junbum submitted a software that was capable of analyzing audio to detect and automatically report crime. While Junbum’s entry was selected as a finalist in the competition, he was less interested in how he placed, and much more interested in answering the questions that arose during the process of creating his submission. “As I was testing, I noticed a lot of limitations and complications. Eventually I thought, ‘Maybe I should just transform this into a scientific paper,” Junbum explains.

That scientific paper, entitled: “Enhancing Reasoning Efficiency and Domain Adaptability of Server-Linked LLMs by Combining with a Customized Deep Learning Model,” was published by the National High School Journal of Science at the end of September. Click the link below to view the full paper:

 Read the full paper here

Inspiration for the paper came when Junbum was attempting to utilize Large Language Models (LLM), but discovered they had limitations, especially when dealing with audio. Junbum decided to test how combining LLM’s with a specialized Deep Learning (DL) model might improve their capabilities. Junbum then ran a series of tests to determine if a DL-enhanced LLM performed better than various baseline LLM’s. The result: Junbum’s new system requires significantly fewer prompt shots for audio classification while reducing average end-to-end processing time by a net improvement of 21.5%.

Beginning in late February, Junbum used his free time to work on his paper, periodically checking in with Avon computer science teacher Evan Sayles for review. He finished it in May and submitted it for publication shortly after. Then, it was subject to the peer review process. “I was most curious about the peer review part. I had never gone through that before, and I was curious to see what others might say or add to my work.”

Junbum doesn’t plan on stopping here. In fact, he has already submitted his next paper, this one exploring how his novel DL-LLM architecture could impact the finance industry. Most notably, Junbum’s idea can predict financial bubbles much cheaper and more accurately than current programs. He says he plans to study business in college, focusing on how it intersects with AI and computer science. “Understanding AI alone won’t lead to making a real-world impact. It’s important to know the business side, too.”

Return To News