Vinith M. Suriyakumar

I am a fifth year PhD student at MIT EECS where I am advised by Dr. Ashia Wilson and Dr. Marzyeh Ghassemi. I also collaborate frequently with Dr. Dylan Hadfield-Menell. I’m supported through a fellowship from Bridgewater Associates and AIA Lab led by Dr. Jas Sekhon where I also conduct research part-time.

Broadly, I’m interested in the privacy, security, and safety of machine learning. The goal of my research is to prevent the misuse of foundation models, especially when it comes to open-sourcing these models. I work on questions related to uncovering risks of open-weight models, applying intepretability to understand these risks, and building safeguards to address these risks.

news

Sep 18, 2025	Our work uncovering and formalizing spurious correlations originating from syntactic shortcuts in LLMs, Learning the Wrong Lessons: Syntatic-Domain Shortcuts in Language Models, was awarded a Spotlight at NeurIPS 2025! This work provides a new perspective on different forms of memorization and uncovers a new type of jailbreak in LLMs.
Jun 3, 2024	Our work on developing a philosophical framework to better reason about the impact of algorithmic decision-making on equal opportunity, Algorithmic Pluralism: A Structural Approach Towards Equal Opportunity, was awarded Best Paper at FAccT 2024!
Jan 15, 2024	Our work on developing an efficient method for estimating privacy with only one training run in federated learning setups, One-Shot Empirical Privacy Estimation for Federated Learning, was awarded an Oral presentation at ICLR 2024!
Apr 29, 2023	Our work on formalizing and analyzing how to fairly use group attributes in prediction models, When Personalization Harms: Reconsidering the Use of Group Attributes in Prediction, was awarded an Oral presentation at ICML 2023!
May 14, 2022	I’m excited to announce that our work from my Summer 2021 internship at Google on Public Data-Assisted Mirror Descent for Private Model Training was accepted to ICML 2022!