Rohan Gumaste

About Me

Welcome to my corner of the internet! I’m Rohan Gumaste, an incoming EECS PhD student at MIT CSAIL, where I intend to work with Professor Michael Carbin.

Recently I graduated from the University of Illinois Urbana-Champaign, where I earned a B.S. in Statistics and Computer Science.

During my undergrad, I worked at the FOCAL Lab with Professor Gagandeep Singh on offline RL, PBRL and adversarial traning for RLHF, as well as applications in AI safety. I was also a part of the ARC lab where I worked with Professor Sasa Misailovic. Here, I researched program synthesis and code generation. Specifically we focused on constrained LLM decoding, with an emphasis on iterative grammar aligned methods. In an earlier project, I briefly worked on translating natural language to formal logic programs.

Earlier I finished a year long remote research experience at Sandia National Laboratories where I worked with Dr. Daniel Ries on statistical modeling for near real-time lightning strike prediction. We studied spatiotemporal modeling, probabalistic programming, bayesian methods, and optimization.

Previously, I worked at the Crowd Dynamics Laboratory under the supervision of Professor Hari Sundaram. We researched deep neural recommender systems, with a focus on fast machine unlearning, differential privacy, and the cold start problem.

Research Interests

Reinforcement Learning: Offline RL, PBRL/RLHF, Robust RL, RLAIF
Programming Languages: Progam Synthesis, Grammar-Guided Code Generation
Machine Learning: Safety, Privacy, Robustness, Generalization Guarantees

News

[May. 2025] Graduated from UIUC
[Jan. 2025] Our paper on Itergen was accepted at ICLR ‘25
[May. 2024] Started as Undergrad RA at UIUC’s ARC Lab
[May. 2023] Started as Undergrad RA at UIUC’s FOCAL Lab
[Aug. 2023] Beginning research internship at Sandia National Labs
[Aug. 2022] Started new position at the Crowd Dyanmics Lab

Publications & Preprints

ICLR

IterGen: Iterative structured LLM generation

Shubham Ugare, Rohan Gumaste, Tarun Suresh, Gagandeep Singh, Sasa Misailovic

The Thirteenth International Conference on Learning Representations (ICLR), 2025.

PDF Code Project Page

Binary Reward Labeling: Bridging Offline Preference and Reward-Based Reinforcement Learning

Yinglun Xu*, David Zhu*, Rohan Gumaste, Gagandeep Singh

PDF Code Under Review.

Universal Black-Box Reward Poisoning Attack against Offline Reinforcement Learning

Yinglun Xu*, Rohan Gumaste*, Gagandeep Singh

PDF Under Review.

Two-Step Offline Preference-Based Reinforcement Learning with Constrained Actions

Yinglun Xu, Tarun Suresh, Rohan Gumaste, David Zhu, Ruirui Li, Zhengyang Wang, Haoming Jiang, Xianfeng Tang, Qingyu Yin, Monica Xiao Cheng, Qi Zeng, Chao Zhang, Gagandeep Singh

PDF Under Review.

Services

Conference Reviewer

ICLR 2025 - The Thirteenth International Conference on Learning Representations