Optimal Autoregressive Chain-of-Thought Learning via Bagging
arXiv preprint, 2026
I’m a PhD candidate in the Department of Statistics & Data Science at Yale University, advised by John Lafferty.
I work on the foundations of artificial intelligence, aiming to identify simple, unifying principles that explain how intelligent systems learn. I’m especially interested in what enables a model to learn complex tasks efficientlyWith as little data, compute, and external supervision as possible. and to generalize beyond its experienceExtending to out-of-distribution problems: longer, harder, or composed in new ways.. I approach these questions as both an empirical scienceExperiments designed to isolate and explain a phenomenon, not just improve a benchmark. and a mathematical scienceFormal theory that explains behavior from statistical and computational principles., developing new architectures, algorithms, and theoretical frameworks.
My work is organized across three interconnected threads:
To learn more, explore my blog, which maps out my work across these interconnected threads.
Select a node on the map to see its details.
What mechanisms an architecture needs for models to learn efficiently and form robust abstractions.
Reasoning in Transformers & LLMs5How models learn, represent, and execute multi-step reasoning, and how to improve out-of-distribution generalization.
Theoretical Foundations6Formal accounts of how architecture, learning algorithm, and problem structure determine what models can learn.
arXiv preprint, 2026
arXiv preprint, 2026
arXiv preprint, 2025
Neural Information Processing Systems (NeurIPS), 2025 Spotlight
International Conference on Machine Learning (ICML), 2025