Foundation models that learn how health unfolds over a lifetime — to see disease coming, reason about it safely, and add healthy years to human lives.
longitudinal health data · temporal modeling · clinical reasoning
Mission and toolkit · click a card to filter papers
recent work — newest first
A generative transformer trained from scratch on longitudinal claims trajectories — diagnoses, procedures, medications and expenditure — scaled up to 1.7B parameters and evaluated on more than 1,000 disease-onset prediction tasks.
Specialized LLM roles with verification gates: generate broad differentials, screen explicitly for must-not-miss conditions and check reasoning against grounded medical evidence.
EEG pre-training that predicts future patches from causal context, with one expert-routing decision per time step shared across all channels to keep them coherent.
A global latent anchor shared by source and target flows gives strong, geometrically stable edits without fine-tuning.
Structured temporal scripts give precise, multi-event timing control over DiT-based video-to-audio generation.
Keeps every visual token at an equal distance from generated text, so the video still shapes long answers.
* equal contribution · Full list on Google Scholar ↗
New preprint: a safety-oriented hypothetico-deductive framework for AI-assisted differential diagnosis.
Released TRACE (EEG representation learning) and foundation models for nationwide medical claims.
Three papers at CVPR 2026 — AnchorFlow (Oral), FoleyDirector (Highlight) and Echoes of Ownership.
Two papers at ICLR 2026: ContextGen and LogiStory.
Two papers at ICCV 2025 and DreamDPO at ICML 2025.
Joined Yale Biomedical Informatics & Data Science as a Postdoctoral Associate.
I'm a Postdoctoral Associate at Yale BIDS, advised by Prof. Hua Xu, building healthcare foundation models with one long-term goal: helping people live longer, healthier lives.
Before Yale I was a postdoctoral researcher at Zhejiang University. I received my Ph.D. from the ReLER Lab, University of Technology Sydney, advised by Prof. Yi Yang, and my M.Eng. and B.Eng. from Xi'an Jiaotong University. My earlier work spans semi-supervised learning, temporal action localization, object tracking, video–language pre-training and controllable generation. Today that multimodal toolkit serves a different question: how can models trained on lifelong health records anticipate disease, guide clinicians, and add healthy years to people's lives?
Postdoctoral Associate, Biomedical Informatics & Data Science · advised by Prof. Hua Xu
Postdoctoral Researcher
Research Intern
Research Intern
Research Intern
Ph.D. in Computer Science, ReLER Lab · advised by Prof. Yi Yang
B.Eng. & M.Eng. in Computer Science · advised by Prof. Deyu Meng
Open to collaboration