Hui Zhang

My long-term research goal is to build language-model agents that learn reliably from interaction, supported by statistically grounded and resource-efficient evaluation. I am particularly interested in the connection between agent learning, feedback quality, and the evidence needed to establish real improvement.

• Continual learning and agent memory — How can agents turn interaction traces into useful experience, transfer it to new tasks, and retain earlier capabilities through updates?
• Reliable reasoning and knowledge use — When should an agent seek evidence, consult another agent, or revise its answer? I study adaptive routing and decision rules that improve reasoning while limiting regressions.
• LLM judges and human feedback — How can evaluation distinguish verifiable errors, stable human preferences, and genuine disagreement? I am interested in judge reliability, preference calibration, and long-horizon creative tasks.
• Efficient statistical evaluation — How can model comparisons use less feedback without losing the information needed for valid conclusions? My interests include exact estimation, uncertainty, coupled sampling, and evaluation during model updates.