Minjun Jo, from 대전대신고등학교 (Daejeon Daeshin High School), became interested not simply in when AI gets an answer wrong, but in what happens when multiple AI agents begin reinforcing the same mistake.
The initial assumption seemed straightforward. If several AI agents examine one another's reasoning, the group should perform better than a single agent working alone.
But after examining test results and conversation logs, Jo observed something more complicated. When several agents produced the same incorrect answer, other agents could begin moving toward that majority view as well.
This raised a simple but important question: Does debate actually make multiple AI systems wiser?
Rather than treating the behavior as an isolated technical failure, Jo focused on understanding why it happened and how the process might be improved. One approach was to introduce a critic agent that reviewed the reasoning generated during debate.
When a Solution Becomes a New Problem
That solution, however, created another problem.
When the critic was used in every situation, accuracy improved, but computational cost also increased significantly. A mechanism designed to improve reliability therefore introduced a new tradeoff between accuracy and efficiency.
For Jo, this was one of the most memorable parts of the research process. Solving one problem did not end the investigation. Instead, the solution itself generated a new question.
That experience shaped the way he now approaches research. Failures and unexpected outcomes are not necessarily setbacks. They can expose constraints that were invisible at the beginning of the study.
From Performance to Explainability
Jo's current interests are focused on AI agents and explainable artificial intelligence, or XAI. Rather than pursuing performance alone, he is interested in systems that can explain why they reached a particular decision and identify potential errors in their own reasoning.
As AI systems become more capable, evaluating how and why they make decisions may become just as important as measuring whether the final answer is correct. Jo hopes to study AI systems that combine strong performance with reliability, interpretability, and the ability to self-check.
Failure as a Starting Point
His experience has also shaped the message he would offer to other student researchers. A hypothesis turning out to be wrong does not mean the research has failed. In some cases, asking "Why did this happen?" after an unexpected result can lead to a more meaningful research question than confirming the original prediction.
For Jo, research is therefore less about avoiding failure and more about refusing to stop at it.
Research Information
Researcher: Minjun Jo (조민준)
Affiliation: 대전대신고등학교 (Daejeon Daeshin High School)