Yoonho Lee

Lee began his Stanford PhD in 2021 after a mathematics bachelor’s and computer-science master’s at POSTECH. His CV records research-scientist work at Kakao and AITRICS and a 2026 student-researcher position at Google Research. His current research asks how to preserve the useful structure of feedback instead of reducing every experience to a single numerical reward.[1][2]

Let the optimizer inspect its full history

Lee is first author of Meta-Harness, which searches over the code controlling an LLM’s information and execution flow. A proposing agent can inspect previous candidates, scores and raw execution traces in a filesystem before making another change. Evaluations and logs accumulate across iterations. This makes detailed failure diagnosis available without placing every past run inside one context window.[4]

Edit an artifact using the reason for a preference

With Joseph Boen and Chelsea Finn, Lee developed Feedback Descent. An evaluator compares a proposed text artifact with the current best and explains its preference. The optimizer uses those critiques to propose a targeted revision. The process changes prompts, code or other textual artifacts at inference time, without updating the underlying model’s weights.[5]

Alternate changes to the model and its harness

Lee coauthored WHALE with Haechan Kim, Gisang Lee, Chelsea Finn and Kangwook Lee. The method alternates model-weight updates with searches for better harness code. Its premise is that improving one component can change which version of the other works best. Experiments examine that interaction on search question answering, mathematical reasoning and chess puzzles.[6]

Retrieve an idea that advances a research question

Lee and Sohyeon Kim are equal contributors to ScholarCatalyst. Its benchmark asks for earlier papers that could advance an initial research question, using judgments and rationales from the authors of completed projects. Restricting candidates to literature available when a project began separates useful prior ideas from hindsight about the finished result.[7]