I’m a Ph.D. candidate at Stanford CS, advised by Chelsea Finn and part of the IRIS lab. I am affiliated with SAIL, CRFM, and the ML Group at Stanford. My research is generously supported through grants and fellowships from OpenAI and KFAS.

I’m developing a new machine learning paradigm where text serves as a primary substrate for storing and updating knowledge. Instead of encoding knowledge solely in neural network weights, I build systems that store and update knowledge directly in text form, modified through text mutations based on rich experiential feedback.

The core vision: enable models to extract massive amounts of information from direct experience (e.g. raw observations, expert feedback, experiment results). As we deploy models on complex, long-horizon tasks, RL’s scalar reward bottleneck will become increasingly limiting. I believe that learning through text can address this by allowing models to learn from a richer set of signals that scale naturally with task complexity.

To this end, I have developed methods for encoding and selecting among a small set of hypotheses about the world [1,2,3] and efficiently fine-tuning model weights [4,5]. I created an interface that enables non-experts to teach vision models via natural language feedback [6]. Most recently, I developed a hierarchical RL framework LLMs discover and leverage textual “abstractions” to solve complex reasoning tasks [7].

My name (윤호) is pronounced like ‘you-know’ said quickly (stress on ‘you’). This is a good approximation.