Hao Li

My core research focuses on developing foundation models that generalize across tasks, modalities, and domains. I explore modular architectures—such as sparse experts, adaptive attention, and continuous-time representations—that can support forecasting, reasoning, and generation across diverse data types including time-series, images, and texts.

I design generative systems that synthesize structured data such as multivariate time series or clinical signals from natural language. My work includes diffusion-based models that allow fine-grained control over the output, enabling precise and faithful generation for simulation, augmentation, and decision support.

I work on building trustworthy AI systems for healthcare, with a particular focus on modeling real-world clinical data such as EHRs, physiological signals (e.g., ECG), and longitudinal records. My research aims to support clinical reasoning, outcome prediction, and health system optimization.

I investigate how large language models perform across tasks, domains, and inputs, with a focus on robustness and attribution. My work includes instance-level analysis of LLM behavior, tracing dataset influence on model predictions, and identifying failure modes in biomedical or instruction-following settings.

I study how representations can be aligned across modalities—such as language, vision, and structured signals—to enable joint reasoning and transfer. My work explores cross-modal distillation, token pruning, and contrastive alignment for both general-purpose LLMs and medical multimodal applications.