I am a fourth-year CS PhD student at Berkeley AI, advised by Emma Pierson and supported by the NSF GRFP.
I work on human-centered AI for science: methods that leverage foundation models to advance how researchers conduct science. I use tools from interpretability to build these methods, as in my recent projects:
- HypotheSAEs, an SAE-based approach to generate hypotheses from text datasets, designed for social science research.
- What’s In My Human Feedback?, a tool for LLM researchers to understand what’s encoded in human feedback data, enabling better post-training.
When applying these methods, I’m drawn to problems involving biomedicine, social science, and the societal impacts of AI.
Previously: I did my undergrad in CS at MIT, and worked with Catherine D’Ignazio, Michael Carbin, and Anshul Kundaje. I’ve interned at Microsoft Research, Apple ML, NVIDIA, and Genesis Therapeutics.
[📧 rmovva@berkeley.edu] [🎓 google scholar] [🐙 rmovva] [🐦 rajivmovva] [👯 my collaborators & mentors]
Recently:
- 🔗 We released a Python package for HypotheSAEs. HypotheSAEs is a method to generate interpretable hypotheses from large text datasets using text embeddings, sparse autoencoders, and LLMs. It’s fast, cheap, and outperforms prompting/clustering/etc. on a variety of datasets (e.g., news headlines, Yelp reviews, Congressional speeches).
Selected Work
What’s In My Human Feedback? Learning Interpretable Descriptions of Preference Data
Rajiv Movva, Smitha Milli, Sewon Min, Emma Pierson.
ICLR 2026
Oral presentation (top 1% of submissions)
[paper] [demo] [code]
Sparse Autoencoders for Hypothesis Generation
Rajiv Movva*, Kenny Peng*, Nikhil Garg, Jon Kleinberg, Emma Pierson.
ICML 2025
[paper] [demo] [code] [pip install] [twitter]
Use Sparse Autoencoders to Discover Unknown Concepts, Not to Act on Known Concepts
Kenny Peng*, Rajiv Movva*, Jon Kleinberg, Emma Pierson, Nikhil Garg.
Preprint
[paper] [twitter]
Annotation alignment: Comparing LLM and human annotations of conversational safety
Rajiv Movva, Pang Wei Koh, Emma Pierson.
EMNLP 2024
[paper] [twitter]
Coarse race data conceals disparities in clinical risk score performance
Rajiv Movva*, Divya Shanmugam*, Kaihua Hou, Priya Pathak, John Guttag, Nikhil Garg, Emma Pierson.
MLHC 2023 (Proceedings) & ML4H 2023 (Findings)
Best findings paper honorable mention at ML4H 2023
[paper] [twitter] [code] [Cornell news] [New York Times]
Topics, Authors, and Institutions in Large Language Model Research: Trends from 17K arXiv Papers
Rajiv Movva*, Sidhika Balachandar*, Kenny Peng*, Gabriel Agostini*, Nikhil Garg, Emma Pierson.
NAACL 2024
[paper] [twitter] [code] [Data Skeptic podcast]
Towards Intersectional, Feminist, Participatory ML: A Case Study in Supporting Feminicide Counterdata Collection
Harini Suresh, Rajiv Movva, Amelia Dogan, Rahul Bhargava, Isadora Cruxên, Ángeles Martinez Cuba, Giulia Taurino, Wonyoung So, Catherine D’Ignazio.
FAccT 2022
Best student paper
[paper] [twitter]
Dissecting Lottery Ticket Transformers: Structural and Behavorial Study of Sparse Neural Machine Translation
Rajiv Movva and Jason Zhao.
BlackboxNLP @ EMNLP 2020
Best paper
[paper] [twitter] [slides] [workshop program]
Deciphering regulatory DNA sequences and noncoding genetic variants using neural network models of massively parallel reporter assays
Rajiv Movva, Peyton Greenside, Georgi K Marinov, Surag Nair, Avanti Shrikumar, Anshul Kundaje.
PLoS ONE, 2019
[paper] [twitter]