I’m a final-year PhD student at Stanford CS, advised by Chelsea Finn. My research focuses on building systems that learn from text directly and continually.

Scalar rewards collapse feedback into a single number. A score gives a verdict, not a diagnosis. What went wrong, why, and what to change is often better expressed in text.

In many settings, we already have access to this kind of feedback: natural-language corrections, pairwise comparisons with explanations, stack traces, and reflections on what worked. Now that models can read text well enough to use it for decision-making, I believe we can do better than learning from scalar rewards alone. My research develops methods that leverage such structured feedback to enable models to continually improve.

The best way to learn more about the technical side of these ideas is to read my blog post or the selected papers below.