I study how agents acquire and learn from new experience while making effective use of what they already know. My focus is on which interactions to pursue, how to interpret feedback, and when prior knowledge should guide behavior or be revised.
Learning can take place through parameter updates, code-policy revision, memory, or skill accumulation. I am interested in a cycle in which new evidence changes subsequent behavior, with improvements that can be retained and reused. This does not require an update after every action.
My work on online reinforcement learning with offline priors examines how deployment evidence should guide the use of existing knowledge. In SCOUT, we investigate how inexpensive exploration can produce experience that helps language agents acquire new capabilities. Together, these questions connect the acquisition of useful experience with the ability to learn from it.