Drew T Nguyen
I am a PhD student in the Department of Statistics at UC Berkeley, advised by Will Fithian and Mike Jordan. I’m also a CBAI research fellow in AI safety.
My current research concerns the measurement and evaluation of AI systems. I’m especially interested in construct validity: what AI evaluations actually measure, in both capabilities and alignment, and how statistical methodology can help clarify their interpretation.
More broadly, I work on statistical methodology and theory, including multiple testing, distribution-free inference, and incentive design. My research has been supported by the NSF Graduate Research Fellowship and the Berkeley Chancellor’s Fellowship.
In the past: I did undergrad at UC San Diego in physics and math, and I’m originally from East San Jose, California—land of burritos and banh mi. Stop by if you’re in town!
You can contact me at drew.t.nguyen[at]berkeley[dot]edu.
Research
On the estimation and validity of AI time horizons—a statistical look at the METR plot.
Drew T. Nguyen and William Fithian. NeurIPS 2026 Workshop on Technical AI Evaluations (poster).[arXiv:2610.12466] [cs.AI], October 2026.
We examine the statistical foundations of METR’s AI time-horizon metric, investigating the relationship between human task duration and AI task difficulty. By fitting splines and using item-response theory, we assess the construct validity of the resulting measurements.
Controlling the false discovery rate under a non-parametric graphical dependence model.
Drew T. Nguyen and William Fithian. Annals of Statistics, forthcoming.[arXiv:2506.24126] [stat.ME], June 2025.
We develop computationally efficient procedures for controlling the false discovery rate under graph-structured dependence. Our methods retain the classical Benjamini–Hochberg procedure’s power when its discoveries form an independent set, while maintaining rigorous error guarantees under dependence.
Incentive Design in Sequential Statistical Protocols Drew T. Nguyen, Alireza Fallah, and Michael I. Jordan. ICML 2026 Workshop on New Frontiers in Game-Theoretic learning / Workshop on Hypothesis Testing, poster.
We study how statistical testing protocols can be designed to incentivize desirable behavior by strategic agents, even when their underlying quality and effort are unobserved.
Data-Adaptive Tradeoffs among Multiple Risks in Distribution-Free Prediction.
Drew T. Nguyen, Reese Pathak, Anastasios N. Angelopoulos, Stephen Bates, and Michael I. Jordan. Neurips 2024 Workshop on Statistical Frontiers in LLMs and Foundation Models (poster).[arXiv:2403.19605] [stat.ME], March 2024.
We develop distribution-free methods for adaptively choosing tradeoffs among multiple statistical risks while maintaining finite-sample guarantees.