Sharan Sahu
I am a second-year PhD student in Statistics and Machine Learning at Cornell University. Broadly, I am interested in statistical machine learning and using statistical/machine learning tools to improve decision-making and bridge the gap between machine learning methodology and practice by making these methods more reliable, human-compatible, statistically rigorous, and deployable. I am fortunate to be supported by a Cornell University Fellowship
Before joining Cornell University, I was an undergraduate at UC Berkeley where I studied Computer Science. I was very fortunate to be advised by Iain Carmichael and Ryan Tibshirani working on statistics and machine learning problems in computational precision medicine and statistical learning theory.
If you're an undergrad or Master's student at Cornell and are interested in collaborating, please reach out!
Email /
Google Scholar /
CV /
Twitter /
GitHub /
LinkedIn
|
|
Research
I am interested in theory and methods in the areas of high-dimensional statistics, stochastic optimization, reinforcement learning, deep learning, language and diffusion models, and differential privacy.
DRO–REBEL: Fast and Robust Policy Optimization for Reinforcement Learning with Human Feedback via Distributional Regression
Sharan Sahu
A family of distributionally robust policy optimization algorithms for RLHF that reduce to simple relative-reward regressions under Wasserstein, Kullback Leibler, and Chi-Squared ambiguity sets, achieving minimax-optimal rates and significantly tighter constants than prior DRO-DPO methods.
[Working Paper]
Towards Optimal Differentially Private Regret Bounds in Linear Markov Decision Processes
Sharan Sahu
Privatizing LSVI-UCB++ with Bernstein bonus achieves state-of-the-art regret in linear MDPs under joint differential privacy with minimal utility loss.
[ArXiv]
|
Presentations and Talks
Towards Optimal Differentially Private Regret Bounds in Linear MDPs
Sharan Sahu
Cornell University, 2025
[Slides]
The Machine Learning Problems Behind Large Language Models: Self-Supervision, Fine-Tuning, and
Reinforcement Learning
Sharan Sahu
University of North Carolina, Chapel Hill, 2025
[Slides]
Unlocking the Power of Databases: The Crucial Role of Theory and Indices in Scalable Vector Databases for Machine Learning
Sharan Sahu
Naval Postgraduate School, 2024
[Slides]
|
Miscellanea
Interview with Aman Manazir: From USAMO Math to Quant Trading, ML PhD, and AI Startups
This is a podcast episode I did with Aman Manazir on the Liftoff Podcast. I share how I first got interested in competitive math and what that journey taught
me about problem-solving. We discuss how I balanced early ML internships and research at Berkeley, Stanford, and USC, the process of deciding between a quant research
role at a top firm and a PhD in machine learning at Cornell, and what it’s like building a small AI startup with LLMs. Along the way we dig into how
to prep for quant trading interviews, land research roles as an undergrad, and take a tech idea from zero to product. Whether you're
curious about ML research, quant trading/research careers, or turning an AI project into a startup, I hope my journey offers a few helpful
takeaways.
Interview with Sithija Manage: My PhD Application Journey
This is an interview I did with Sithija Manage about my journey
through graduate school applications. We discuss my undergraduate research journey at Berkeley, Stanford, and USC, and how those early
experiences guided my focus toward a PhD. I detail the graduate admissions process: writing a clear and impactful statement of purpose,
building meaningful mentorships that led to supportive recommendation letters, departmental visits, and what to look
for in a graduate program that fits your career and research goals. Along the way, I try to offer some advice on securing undergrad
research roles, crafting a cohesive application narrative, and evaluating program decisions. If you're considering doing a PhD or
interested in getting into research, I hope this video offers some helpful advice.
|
|