I am Rachit, a PhD student at Harvard University advised by Prof. Sham Kakade and Prof. David Alvarez-Melis.

Over the past few years, I walked my first baby steps as a researcher owing to some wonderful people and collaborations. Most recently, I was a pre-doctoral researcher at Google DeepMind, working on modularizing LLMs with Partha and Prateek. Before that, I pursued my bachelor’s thesis research with Yonatan at the Technion in Israel studying the information-theoretic properties of neural networks. Before that, I was a research intern at Adobe’s Media and Data Science Research Lab, where I worked on commonsense reasoning for large language models.

I was fortunate to collaborate with Danish on evaluating neural attribution methods¹. I also had an amazing time working with Naomi studying mode connectivity in loss surfaces of language models².

I also spent a couple of wonderful summers as a part of the Google Summer of Code program with the Cuneiform Digital Library Initiative (CDLI). Here, I was advised by Jacob and Niko.

News and Timeline

2026

• January Our work at Meta: Test-time Training for long-context LLMs and RL Scaling Laws accepted at ICLR 2026!

2025

• May Interning at Meta Superintelligence Lab (MSL) this Summer!

2024

• August Starting my doctorate at Harvard University!
• May Presenting our work on composing LLMs at ICLR 2024 in Vienna!

2023

• May Presenting our work on linear mode connectivity at ICLR 2023 in Kigali! With Jeevesh and Naomi.

2022

• September My bachelor’s thesis work done at the Technion was accepted at NeurIPS 2022!
• August Joining Google DeepMind as a pre-doctoral researcher!
• May Two papers on commonsense and factual reasoning done at Adobe MDSR accepted at NAACL 2022!
• January Starting my bachelor’s thesis with Yonatan at the Technion, Israel!

2021

• November Our work evaluating model explanations is accepted at TACL! In collaboration with Danish and others at CMU. <!– * May Work with CDLI accepted at ACL SRW 2021. Gauging machine translation and sequence labeling for extremely low-resource languages.
• May Starting as a Research Intern at Adobe’s Media and Data Science Research (MDSR). –>