Atharva Kulkarni

Hello! I am a 3rd year CS PhD student at the University of Southern California. I am advised by Swabha Swayamdipta and am member of DILL Lab and USC NLP Group.

I study the science of foundation models – why they succeed, why they fail, and how to make them more robust and reliable.

Before joining USC, I completed my Masters in Language Technologies (MLT) from Carnegie Mellon University – Language Technologies Institute, where I worked with Barnabás Póczos & Graham Neubig. I was also fortunate to collaborate with Aditi Raghunathan and Ameet Talwalkar.

Prior to CMU, I was a Predoctoral Researcher at the Laboratory for Computational Social Systems (LCS2), IIT Delhi. I graduated from Savitribai Phule Pune University with a Bachelor’s degree in Computer Science.

I have also spent time at Apple, Microsoft, Thoughtworks, and UC Berkeley – Simons Institute for the Theory of Computing.

Research

My broad research focus is on theoretical & empirical understanding of foundation models. Specifically, I am interested in:

Building a principled understanding of when & why they work / fail.
Studying their learning dynamics, geometric properties, & structural constraints.
Exploring avenues for improving their reliability, safety, & trustworthiness.

These days I am interested in how the design, data, and optimization choices of foundation models affect their performance and representation geometry. My goal is to understand why certain approaches work / fail in practice and how they can be improved.

I’m eager to connect with my academic peers! If our research interests align (or diverge) in intriguing ways, I’d be delighted to explore potential collaborations or simply exchange ideas!

News

May 2026	Started my summer research internship at Microsoft Research with the Teams Core Applied Science group!
Apr 2026	Disentangling Geometry, Performance, and Training in Language Models is accepted at ICML 2026 as a spotlight paper! See you in Seoul !
Jan 2026	Joining Thoughtworks AI Labs as a Student Researcher for Spring 2026!

🕰️ all news ...

Selected Publications

ICML

Disentangling Geometry, Performance, and Training in Language Models

Atharva Kulkarni, Jacob Mitchell Springer, Arjun Subramonian, and Swabha Swayamdipta

In Forty-third International Conference on Machine Learning, Jul 2026

Spotlight Abs

Geometric properties of Transformer weights, particularly the unembedding matrix, have been widely useful in language model interpretability research. Yet, their utility for estimating downstream performance remains unclear. In this work, we systematically investigate the relationship between model performance and the unembedding matrix geometry, particularly its effective rank. Our experiments, involving a suite of 108 OLMo-style language models trained under controlled variation, reveal several key findings. While the best-performing models often exhibit a high effective rank, this trend is not universal across tasks and training setups. Contrary to prior work, we find that low effective rank does not cause late-stage performance degradation in small models, but instead co-occurs with it; we find adversarial cases where low-rank models do not exhibit saturation. Moreover, we show that effective rank is strongly influenced by pre-training hyperparameters, such as batch size and weight decay, which in-turn affect the model’s performance. Lastly, extending our analysis to other geometric metrics and final-layer representation, we find that these metrics are largely aligned, but none can reliably predict downstream performance. Overall, our findings suggest that the model’s geometry, as captured by existing metrics, primarily reflects training choices rather than performance.
TMLR

Multitask Learning Can Improve Worst-Group Outcomes

Atharva Kulkarni^*, Lucio M. Dery^*, Amrith Setlur, Aditi Raghunathan, Ameet Talwalkar, and Graham Neubig

Transactions on Machine Learning Research, Feb 2024

Abs Code

In order to create machine learning systems that serve a variety of users well, it is vital to not only achieve high average performance but also ensure equitable outcomes across diverse groups. However, most machine learning methods are designed to improve a model’s average performance on a chosen end task without consideration for their impact on worst group error. Multitask learning (MTL) is one such widely used technique. In this paper, we seek not only to understand the impact of MTL on worst-group accuracy but also to explore its potential as a tool to address the challenge of group-wise fairness. We primarily consider the standard setting of fine-tuning a pre-trained model, where, following recent work (Gururangan et al., 2020; Dery et al., 2023), we multitask the end task with the pre-training objective constructed from the end task data itself. In settings with few or no group annotations, we find that multitasking often, but not consistently, achieves better worst-group accuracy than Just-Train-Twice (JTT; Liu et al. (2021)) – a representative distributionally robust optimization (DRO) method. Leveraging insights from synthetic data experiments, we propose to modify standard MTL by regularizing the joint multitask representation space. We run a large number of fine-tuning experiments across computer vision and natural language processing datasets and find that our regularized MTL approach consistently outperforms JTT on both average and worst-group outcomes. Our official code can be found here: https://github.com/athrvkk/MTL-group-robustness.