Research

My interests are in statistically efficient learning and in how geometric structure shapes training dynamics and generalization in machine learning. In particular, I am interested in the fundamental limits of learning and algorithms that reach them, as well as theoretical foundations of modern machine learning methods.

I work on statistical learning theory with Nikita Zhivotovskiy in Berkeley's Statistical Machine Learning group, and on deep learning theory with Michael DeWeese at BAIR.

I also work on computational nuclear physics at LLNL.

Google Scholar: link. ORCID: 0009-0004-1252-1679.

(Short writeups of some of this work are available here.)

Publications

(\(\alpha\beta\)) denotes alphabetical ordering, * denotes equal contribution.

Majority-of-Three is Optimal

Divit Rawal, N. Zhivotovskiy (\(\alpha\beta\))

Preprint. Under review. [arXiv] / (tldr)

A Theory of Saddle Escape in Deep Nonlinear Networks

Divit Rawal, M. R. DeWeese

Preprint. Under review. [arXiv] / (tldr)

Rao-Blackwellized Score Matching on Manifolds

Divit Rawal

ICML 2026, Workshop on Structured Probabilistic Inference and Generative Modeling. [arXiv] / [poster] / (tldr)

Minimax Rates for Hyperbolic Hierarchical Learning

Divit Rawal, S. Vishwanath

Preprint. [arXiv]

ALPHANSO: Open-Source Modeling of (\(\alpha\),n) Neutron Source Terms

Divit Rawal, A. J. Nelson, W. Zywiec, D. Siefman

Nuclear Instruments and Methods in Physics Research Section A. [doi]

American Nuclear Society Student Conference, 2026. (Best Paper in Mathematics, Computation, and AI Applications). [poster]

INMM Annual Meeting 2026.

Invited talks at LBNL, LLNL, UT Knoxville.

I have also contributed to the development and releases of Foundation-Sec-8B-Instruct and Foundation-Sec-8B during an internship at Cisco Foundation AI, and worked on two ATLAS analyses using neural simulation-based inference, for off-shell Higgs production and for parameter estimation, as a researcher in the Whiteson lab.