I am a postdoctoral scholar at the Halıcıoğlu Data Science Institute, UC San Diego, working with Misha Belkin and Dima Drusvyatskiy.
My research aims to advance the scientific understanding of modern machine learning by leveraging small-scale settings to build faithful abstractions of phenomena seen in practice. Recently, I’ve been interested in how the inductive biases arising from the optimization algorithm, architecture, and training data and objectives shape learning dynamics and generalization. I am also interested in how LLMs represent concepts and use them to adapt to new tasks.
I completed my PhD in Computer Science at USC, advised by Vatsal Sharan of the USC CS Theory Group, working closely with Christos Thrampoulidis and Mahdi Soltanolkotabi. Before that, I graduated from IIT Roorkee with a bachelor’s in ECE, where I worked on generative adversarial network-based methods for compressive sensing MRI reconstruction.
News
- Aug 2026Started as a postdoc at HDSI, UCSD.
- May 2026Passed my Thesis Defense at USC.
- May 2026Co-organizing MOSS workshop at COLM 2026.
- April 2026Selected for the USC Viterbi Best Research Assistant Award (one per dept).
- March 2026Invited Talk at CPAL 2026 as a Rising Star.
Publications
Preprints
Understanding Contextual Recall in Transformers: How Finetuning Enables In-Context Reasoning over Pretraining Knowledge
2026
In-Context Benign Overfitting: A Feature-Selection Model in In-Context Linear Regression
NeurIPS 2026
2025
In-Context Occam’s Razor: How Transformers Prefer Simpler Hypotheses on the Fly
COLM 2025; MOSS Workshop at ICML 2025 (Oral)