I am a final-year PhD student in Computer Science at Stanford University, advised by Chelsea Finn. My research centers on continual learning in LLMs and is supported by OpenAI and KFAS.

Genuine continual learning is more than quick local adaptation; it means improving over time by learning from a wide range of experiences based on history. Text is a natural medium for this learning. Many forms of textual feedback contain useful information on how to improve a system: execution traces, scored or annotated trajectories, human critiques, retrieved articles, and experiment logs all provide insight into what went wrong and what to try next. Standard learning pipelines compress this information too early, reducing it to a scalar objective for gradient computation. External textual artifacts offer a natural place to store and reason over such information without discarding structure.

I am working on three main research areas, all focused on making true continual learning possible:

1. Better text optimizers. Improving text artifacts for a specific goal is key to continual learning with text. Recently, I introduced Meta-Harness, a method that saves all previous information in a filesystem; this enables long-term credit assignment without enlarging the context window.

2. Deciding what to optimize. Models are improving at optimizing almost any reasonable fixed goal on their own, and advances in text optimization will speed this further. The main challenge now is choosing what to optimize. I am interested in learning systems where both the target and the goal change together: new artifacts reveal gaps in the current goal, and new goals guide the next search.

3. Understanding text-space learning. Learning directly from text works well, but we still do not know much about how or why it works. I want to develop a clear framework for understanding the “text hypothesis class” and the algorithms that use it.

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