Ziran Yang

I view exploration as the core challenge in RL, and test-time search being necessary to achieve it for LLM agents. The key technical problem is how search procedures and expert decision-making systems can be internalized as reasoning ability, rather than remaining external scaffolding. I see two tightly coupled aspects: backfilling expert search behavior into the model through learning, and on-the-fly calibration that lets the model assess uncertainty and decide when to search, explore, or trust its own prediction.

Email Scholar X GitHub