Ziran Yang

I view exploration as a central challenge in reinforcement learning, with test-time search serving as a natural mechanism for exploration in LLM agents. My research asks how models can internalize the benefits of search and external feedback, rather than rely on them solely as external scaffolding. I study two tightly coupled directions: distilling expert search behavior into the model through learning, and calibrating the model at test time to recognize uncertainty and decide when to search, explore, or rely on its own predictions.

Email Scholar X GitHub