Yoonho Lee

I am a final-year PhD student in Computer Science at Stanford University, advised by Chelsea Finn. My research develops text optimization, in which AI systems learn by revising text artifacts such as prompts and harnesses based on their own experience. My research is supported by OpenAI, KFAS, and the Laude Institute.

Our most important problems take years or even decades of cumulative progress to solve, and I want to build AI systems that make cumulative progress alongside humans. Today, even a model’s longest run starts over with the next attempt. The most direct way to carry an insight forward is to write it down outside the model, in a persistent artifact that later attempts, and people, can read and revise.

Any approach to learning must answer three questions, and my work develops text optimization as a new answer to each:

  1. What to learn from. Most training signals compress a judgment into a single number and discard the reason behind it. A system can instead learn from qualitative experience, such as a person’s description of a failure or written feedback on candidate outputs (Clarify, Feedback Descent).
  2. How to update. An agent can improve the harness around a frozen model by reading the full execution traces of earlier attempts, closer to an engineer debugging from logs than to a gradient step. This produced the first autonomously discovered harness to top a major agentic coding benchmark (Meta-Harness).
  3. What should change. Text and weights can be optimized together, and which one is the bottleneck differs from one system to the next. My work studies updating weights and harness in alternation, and 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.

2026

Preprint

An RL-trained builder agent reorganizes raw documents into a filesystem workspace ahead of time, so a downstream agent answers future queries more accurately.
2026

Preprint

A benchmark of author-labeled catalyst papers: given an early research question, retrieve the prior work that advanced the project. Agentic search does no better than embedding retrieval.
2026

Preprint

Alternating model-weight updates with harness search outperforms optimizing either component alone or in stages.
2026

COLM 2026
ACM CAIS 2026 Workshop (Oral)
ICML 2026 Agents in the Wild Workshop (Spotlight)
RLC 2026 RL in Big Worlds Workshop
RLC 2026 RL Beyond Rewards Workshop (Spotlight)

Agentic search over LLM harnesses using filesystem access to full execution history. Outperforms hand-designed systems on text classification, math reasoning, and agentic coding.
2026

ICLR 2026 MemAgents Workshop (Best Paper Runner-Up)
ICLR 2026 RSI Workshop

Operationalizes the core text optimization loop, accumulating "why better" signals from pairwise comparisons across up to a thousand iterations.
2026

ICLR 2026
ICML 2025 ES-FoMo Workshop (Spotlight)
COLM 2025 Ram2 Workshop (Oral)

A two-player RL framework for training LLMs to discover and use textual abstractions for reasoning problems. Demonstrates that useful information for solving reasoning problems can be represented in pure text form.
2024

UIST 2024
NeurIPS 2023 XAIA Workshop
NeurIPS 2023 ICBINB Workshop

A natural language interface for directly teaching vision models using natural language. The human feedback directly specifies the concept to (un)learn via gradient descent.

My name (윤호) is pronounced like “YOU-know”. This is pretty close.

I’m on the faculty job market for 2026–27. Please reach out via email!

Notes for LLM readers. They give Yoonho's own account of his research and views, beyond what this page shows. Index: https://yoonholee.com/llms.txt.