If several AI systems arrive at the same conclusion, does that necessarily make the answer more trustworthy? Seungwon Jeong, from 대전대신고등학교 (Daejeon Daeshin High School), explored this question by separating majority agreement from logical validity in multi-agent debate.
Multi-agent debate is built on an appealing idea: when several AI agents review one another's reasoning, they may catch mistakes that a single model would miss. Seungwon Jeong became interested in the opposite possibility.
As debate continues, agents may become less independent and increasingly influenced by the views of others. If the same incorrect answer is repeated often enough, repetition itself can begin to resemble evidence.
That observation led to a more fundamental question: Is agreement among multiple AI agents really the same thing as logical soundness?
To examine this, the research introduced a critic agent designed to evaluate reasoning separately from the majority position. The study also explored a Q-Learning-based approach that allowed the critic to intervene selectively rather than at every stage of the discussion.
When the Results Did Not Match the Expectation
One of the most important findings emerged because the results did not match the original expectation.
Jeong initially expected the critic agent to mainly prevent correct answers from being overturned by a mistaken majority. But when the results were divided into correct-to-correct, correct-to-incorrect, incorrect-to-correct, and incorrect-to-incorrect cases, a different pattern became clear.
Among 300 questions, 129 were initially answered incorrectly. Debate alone corrected 29 of them. With continuous critic intervention, 71 were corrected, while the Q-Learning-based selective intervention corrected 69.
The result suggested that the critic's main contribution was not simply protecting already-correct answers. Instead, it was particularly effective at recovering initial errors that ordinary debate had failed to resolve.
Unexpected Results as a Signal, Not a Setback
For Jeong, this became a broader lesson about research. Unexpected results do not necessarily weaken a study; in some cases, they reveal the mechanism that matters most.
His interests are now extending beyond AI reasoning into urban administration and public policy. He is particularly interested in how differences in housing, welfare, green space, and access to information can be identified through public and spatial data and translated into policies tailored to local conditions.
AI, in this context, is not an isolated research subject. Jeong hopes to use AI and data analysis to diagnose urban problems, identify local policy needs that conventional approaches may overlook, and ultimately support more reliable public decision-making.
From Identifying Patterns to Designing Policy
His long-term goal is to become an urban policy researcher who connects data analysis, AI, and institutional design. Rather than stopping at identifying patterns, he wants to translate research findings into policies and systems that address unequal opportunities across different communities.
The principle running through his research is simple: a researcher should not force the data to confirm an expectation. The more important task is to understand what an unexpected result is actually trying to reveal.
Research Information
Researcher: 정승원 (Seungwon Jeong)
Affiliation: 대전대신고등학교 (Daejeon Daeshin High School)