I am broadly interested in understanding deep learning.
When do models generalize, and what should generalization mean?
When models fail, do they fail in predictable ways?
I like to identify an interesting empirical behavior
(e.g. length-generalization, or calibration, or...),
and then characterize the behavior as precisely as possible.
Separately, I've recently been thinking about new & improved mechanisms for peer review.
Interns I've Hosted/Collaborated
Faculty I've Sponsored (via Apple)
Service
Area Chair: ICLR 2024, NeurIPS 2024, ICML 2025, ICML 2026
Research
Calibration
Calibration is one way of formalizing how models which are not Bayes-optimal can still be "good" in other ways.
What types of calibration hold for deep networks, and why?
Is there a meaningful and achievable notion of calibration for LLMs?
-
Trained on Tokens, Calibrated on Concepts: Semantic Calibration in LLMs
2025
arXiv
Preetum Nakkiran, Arwen Bradley, Adam Goliński, Eugene Ndiaye, Michael Kirchhof, Sinead Williamson
-
When Does Optimizing a Proper Loss Yield Calibration?
NeurIPS 2023 Spotlight
arXiv
Jaroslaw Blasiok, Parikshit Gopalan, Lunjia Hu, Preetum Nakkiran
-
A Unifying Theory of Distance from Calibration
STOC 2023
arXiv
code
Jarosław Błasiok, Parikshit Gopalan, Lunjia Hu, Preetum Nakkiran
Diffusion & Generative Models
I've written a tutorial on diffusion models,
studied the mechanisms of diffusion generalization (classifier-free guidance, composition),
and helped develop new methods (TarFlow).
-
Normalizing Flows are Capable Generative Models
ICML 2025 Oral
arXiv
Shuangfei Zhai, Ruixiang Zhang, Preetum Nakkiran, David Berthelot, Jiatao Gu, Huangjie Zheng, Tianrong Chen, Miguel Angel Bautista, Navdeep Jaitly, Josh Susskind
-
Step-by-Step Diffusion: An Elementary Tutorial
Foundations and Trends 2024
arXiv
book
Preetum Nakkiran, Arwen Bradley, Hattie Zhou, Madhu Advani
Understanding Generalization
Why do overparameterized models generalize?
Why do underparameterized models generalize?
Are these questions related?
Why do LLMs generalize out-of-distribution, and how do we formalize this?
-
What Algorithms can Transformers Learn?
ICLR 2024
arXiv
Hattie Zhou, Arwen Bradley, Etai Littwin, Noam Razin, Omid Saremi, Josh Susskind, Samy Bengio, Preetum Nakkiran
-
The Deep Bootstrap Framework
ICLR 2021
arXiv
Preetum Nakkiran, Behnam Neyshabur, Hanie Sedghi
-
Deep Double Descent
ICLR 2020
arXiv
Preetum Nakkiran, Gal Kaplun, Yamini Bansal, Tristan Yang, Boaz Barak, Ilya Sutskever
-
Towards an Empirical Theory of Deep Learning
PhD Thesis 2021
pdf
Preetum Nakkiran
About
I completed my PhD at Harvard, advised by
Madhu Sudan and
Boaz Barak.
In my postdoc I worked with Misha Belkin.
I did my undergrad in EECS at UC Berkeley. Go Bears!
I'm broadly interested in theory and science. In the past, I've interned at
OpenAI (with Ilya Sutskever), Google Brain (with Behnam Neyshabur, Hanie Sedghi),
and have done research in error-correcting codes, distributed storage, and cryptography.
I'm grateful for past support from NSF GRFP, the Google PhD Fellowship, and the Simons Institute.
For talks, use this bio.
An (outdated) CV is here.
ORCID
What People are Saying
a "high-level" scientist
— colleague (ML)
makes plots and draws lines through them
— colleague (TCS)
has merits that outweigh flaws
— reviewer 2
From The Archives
- Past successful application materials: drive
- Courses I took in undergrad