Swaroop Nath

Swaroop Nath

Pre-Doctoral Researcher @ Google DeepMind

Experience

Pre-Doctoral Researcher

Google DeepMind

July 2024 - Present

AI/ML Research Intern

LinkedIn

May 2023 - July 2023

Education

MS by Research, CSE

Indian Institute of Technology, Bombay

Aug 2021 - June 2024

About Me

Hi! I am a Pre-Doctoral Researcher in Google DeepMind India. Here I work with the amazing folks in the Machine Learning and Optimization team, led by Dr. Prateek Jain. My work mostly revolves around efficiency — where I initially dabbled with Efficient Attention, and eventually moved to Efficient Reasoning.

Prior to this, I graduated as an MS by Research scholar from the CFILT Lab, IIT Bombay. I was fortunate enough to be advised by Dr. Pushpak Bhattacharyya (God rest his soul) and Dr. Harshad Khadilkar. My Thesis was recognized with the Winifred B. Fernandes Research Excellence Award. My works have been featured in A* NLP conferences — EMNLP and ACL.

News

Publications & Pre-Prints (* = Equal Contribution)

Leveraging Domain Knowledge for Efficient Reward Modelling in RLHF

Swaroop Nath*, Tejpalsingh Siledar*, Pushpak Bhattacharyya, et al.

arXiv Preprint

This research explores a novel reward modeling approach for Reinforcement Learning from Human Feedback (RLHF) that significantly reduces the need for human preference data by leveraging domain knowledge, demonstrated in the context of e-commerce opinion summarization.

One Prompt To Rule Them All: LLMs for Opinion Summary Evaluation

Swaroop Nath*, Tejpalsingh Siledar*, Pushpak Bhattacharyya, et al.

ACL 2024

This work introduces SUMMEVAL-OP, a benchmark dataset for evaluating opinion summaries, and proposes two novel prompting strategies (OP-I-PROMPT and OP-PROMPTS) that significantly improve the correlation of LLM-based evaluations with human judgments, especially for open-source models.

Transformers are Expressive, But Are They Expressive Enough for Regression?

Swaroop Nath, Harshad Khadilkar, and Pushpak Bhattacharyya

arXiv Preprint

This work investigates the theoretical and empirical limitations of Transformer models in approximating smooth functions, a fundamental aspect of regression tasks. The findings suggest inherent constraints on their expressivity in this domain.

Reinforcement Replaces Supervision: Query focused Summarization using Deep Reinforcement Learning

Swaroop Nath, Pushpak Bhattacharyya, and Harshad Khadilcar

EMNLP 2023

This paper proposes a novel reinforcement learning (RL) algorithm for query-focused summarization (QfS), achieving significant improvements over state-of-the-art supervised methods. The approach also introduces a new passage embedding scheme to create a more effective reward function for training.