Papers & projects
Statistical learning and large language models.
Preprint · 2026
OPD provides dense token-level supervision. We formulate optimal OPD as bilevel optimization, learning how to weight teacher signals to improve the student’s performance.
Preprint · 2026
Also accepted at NeurIPS 2026 workshops: MLxOR · RAAAI · TTCL
We derive a closed-form asymptotic cost–accuracy Pareto frontier for LLM test-time compute, then use it to guide adaptive rollouts and voting under a token budget.
The Annals of Statistics · 2026 · 54(2), 570–596
A leading journal in statistical theory and methodology.
We identify the model that maximizes worst-case predictive performance over mixtures of source populations, with no target labels required.
Preprint · 2025
Learning shared representations by maximizing worst-case explained variance across heterogeneous data sources.
Preprint · 2025
Developing robust classifiers and statistical inference for multi-source domain adaptation with unlabeled target data.
Preprint · 2024
Turning causal invariance learning into a continuous nonconvex optimization problem with statistical and computational guarantees.
The R Journal · 2024
Providing practical tools for confidence intervals and hypothesis testing in high-dimensional linear and logistic regression.