Research

My work turns auxiliary information into calibrated evidence. I use historical and negative controls to expose residual confounding, and flexible machine-learning methods to learn heterogeneous treatment effects and identify sharp sensitivity bounds when standard identifying assumptions may fail. More recently, I have extended this perspective to combining abundant but noisy auxiliary information with limited high-quality validation data.

The unifying goal is to make hidden bias and imperfect labels empirically diagnosable, correctable, and useful for inference—rather than simply assuming them away.

I study how to retain the inferential value of pooled data when observations are distributed across institutions, heterogeneous across populations, or missing in incompatible blocks. My work first showed that more aggregation can hurt: the gain from larger samples must be balanced against bias from heterogeneity, and communication structure determines which sites should borrow from one another.

My recent work shifts the focus from aggregating model parameters to sharing information-preserving statistical objects. Transfer functions can safely augment an internal study with blockwise-missing external data; distributed calibration can reproduce pooled-data treatment comparisons; and MOSAiC compresses entire local risk functions so that sites communicate once while supporting accurate, flexible downstream analyses.

I am interested in using rigorous statistical tools, such as non-asymptotic theory and optimal testing, to study the full lifecycle of modern learning systems. On the learning side, I investigate why overparameterized neural networks generalize and how continual-learning algorithms balance forward transfer against retention of earlier knowledge. On the verification side, our team formulates LLM watermarking as a statistical testing problem and derives detection rules with explicit efficiency and optimality guarantees.

A recurring theme is that familiar methods can fail in modern regimes: classical bias–variance intuition can miss the benefit of overparameterization, standard regularizers can be suboptimal across tasks, and additive watermark detectors lose robustness when humans edit generated text. The theory reveals these failure mechanisms and points to better procedures.

I develop methods that use geometry, sparsity, and relational dynamics to reveal interactions hidden by nonstandard observations. CARE shows that although compositional measurements obscure absolute scale, sparsity and increasing dimension can restore identifiability: in sufficiently high dimensions, its precision-matrix estimator is minimax optimal and performs as if the latent basis were observed. Related work recovers nonlinear dependence graphs from asynchronous event streams and models out-of-distribution fluid dynamics through disentangled graph ODEs.

This direction treats structure not as a complication to work around, but as information that can make otherwise difficult learning and inference problems solvable.