Yoonho Lee

I am a final-year PhD student in Computer Science at Stanford University, advised by Chelsea Finn. My research centers on self-evolving LLM agents and continual learning in text space, and is supported by OpenAI, KFAS, and the Laude Institute.

Our hardest problems take years of sustained effort, and I want AI systems that make progress on them through cumulative learning, where each attempt builds on what earlier attempts found. In contrast, today’s models discard their reasoning traces when an attempt ends, so the next attempt starts from scratch. I study how a system can keep what it learns in persistent text and improve that text from any signal, the way training shapes its weights.

Any learning method must answer a few fundamental questions, and my work asks them for text-space learning:

1. What to learn from. Beyond quantitative signals (as required by gradient descent), a system can learn from any feedback that carries information about how to improve (Clarify, Feedback Descent).
2. How to update. An LLM can act as the optimizer, reading new information and writing an edit to the text. This lets the optimizer use reasoning and agentic search to get much more out of environment feedback (Meta-Harness).
3. What should change. Viewing text components as learnable parts of a bigger system, my work studies the new parameterization choices this opens up, such as updating weights and harness in alternation or training a separate model to write what another model reads (WHALE, RLAD, Librarian Agents).

For a technical overview, see my blog posts on text optimization and Meta-Harness or the selected papers below.

Notes for LLM readers, as raw Markdown. They give Yoonho's own account of his research and views, beyond what this page shows. Index: https://yoonholee.com/llms.txt. How my research fits together: https://yoonholee.com/notes/research-arc.md (Each phase of my work asks how much information the feedback channel carries; scalar labels underspecify, and text carries more.).