Zonghai Yao

I am a Ph.D. candidate in Computer Science at the University of Massachusetts Amherst (expected May 2026), working with Prof. Hong Yu in the UMass BioNLP Lab. I work at the intersection of AI, NLP, and health. My long-term goal is an “AI Hospital”: a set of reliable AI agents that can think and also communicate across real clinical workflows. In this vision, agents are not standalone chatbots; they are tool-using teammates that can reason over evolving patient context, coordinate across roles, and stay verifiably grounded in clinical evidence. I am especially excited about promising directions like long-horizon agent learning in realistic simulations, multimodal clinical world models, and evaluation that ties model behavior to safety- and outcome-relevant signals. Previously, I received my B.S. in Computer Science from Nankai University.

My work focuses on three connected problems:

1. A central focus is agents that think in high-stakes settings. I build LLM agents for evidence-based clinical reasoning, where models plan, retrieve guidelines or knowledge graphs, and justify decisions with verifiable support rather than fluent guesses. I am also deeply interested in multimodal medical agents in “multiple images” settings—where “multiple” can mean time (longitudinal imaging across a patient’s care journey, or video such as surgical procedures), space (3D volumes), and modality (e.g., X-ray, CT, MRI, polysomnography). My goal is to develop agents that can reason over these heterogeneous streams in a unified, step-by-step way.

2. I also study agents that communicate with patients and clinicians over long horizons. I build multi-agent simulations for discharge education, chronic disease management, and recovery support, with a focus on multi-session interactions where goals unfold over days, weeks or months. I evaluate whether agents can maintain state, adapt explanations to health literacy and emotion, and remain safe and helpful under realistic conversational drift and social pressure.

3. I develop methods to optimize, train, and stress-test these agents. On the system side, I design test-time orchestration loops (planning, retrieval, tool use, self-check) that make reasoning more structured and auditable. On the learning side, I use data-centric pipelines (expert annotation, human edits, synthetic data, hard-case mining) and learning from feedback (preference signals, outcome-verifiable objectives, and RL-style training in simulation) to improve faithfulness, robustness, and long-horizon behavior. Across projects, I pair these with fine-grained evaluation that diagnoses failures at the step level and under distribution shift, so we can iterate on both models and agent policies.

Service