Events

The Division of Health Informatics and Artificial Intelligence is excited to host Dr. Ruihao Zhu, Assistant Professor at the Cornell SC Johnson College of Business.

AI systems are beginning to change two core steps of knowledge discovery: gathering evidence and running experiments. This talk presents two complementary lines of work toward making these processes more reliable and sample-efficient.

In the first part, Zhu will discuss deep research agents: LLM-based systems that use external tools such as web search and web reading to answer complex factual queries. Zhu formalizes deep research as an episodic Markov decision process and uses this formulation to isolate which design choices actually improve reinforcement-learning-based agents. The key findings are that AI feedback from an LLM judge, on-policy RLOO training, careful data curation, and test-time error tolerance substantially improve performance. Somewhat surprisingly, several intuitive choices, such as format rewards and multi-epoch GRPO updates, do not help or can even hurt.

In the second part, Zhu turns to accelerating experimentation through generator-augmented best-arm identification. In settings such as clinical trials or policy selection, collecting a real sample can be costly, while LLMs or predictive models can generate synthetic outcomes at low cost. Notice that directly adopting synthetic outcomes can introduce bias; Zhu thus uses them as centered control variates to reduce estimation variance. The resulting algorithm can provably reduce sample complexity while preserving statistical guarantees.

Register to attend