Mayank Shrivastava
PhD CS, UIUC'28
I am a second-year Ph.D. student in Computer Science at the University of Illinois at Urbana–Champaign, advised by Prof. Arindam Banerjee.
Research Interests
My research focuses on generative modeling, with emphasis on both theory and applications, particularly in controlled generation and data assimilation. I am also broadly interested in Reinforcement Learning and Federated Learning.
Previous Experience
- Amazon Applied Science – Research Intern with John McKay (Summer 2025)
- National University of Singapore – Research Intern under Prof. Jonathan Scarlett (Summer 2022)
- Samsung Electronics, Suwon – Machine Learning Engineer, Samung Bixby(2020–2021)
- IIT Kanpur, India – B.Tech, Electrical Engineering (2016–2020)
Checkout my Incomplete notes.
Click here for CV.
Publications
2026
- Under submission
TL;DR
We propose Flow Proposal Particle Filters (FPPF), which use a conditional generative model to learn near-optimal proposals for particle filters. By conditioning particle propagation on observations, FPPF steers particles toward high-posterior regions before weighting, reducing weight variance and mitigating particle degeneracy in high-dimensional nonlinear systems, while retaining exact Bayesian filtering via tractable likelihoods.
- Under submission
TL;DR
In this work, we study stochastic optimal control formulations of generative modeling, where sampling is framed as learning a controlled stochastic process. We show that Adjoint Matching performs descent on a surrogate objective and can become unstable due to uncontrolled distribution shift between path distributions. To address this, we propose Fisher Adjoint Matching (FAM), a natural-gradient method that approximates a KL trust-region update, yielding more stable and efficient training of generative models.
2025
- AISTATS (Spotlight)
TL;DR
In this work, we develop a general framework to obtain **uniform bounds for sketched bilinear forms** using geometric complexity and generic chaining. Our results **unify and extend classical guarantees such as the Johnson–Lindenstrauss lemma and RIP**, and lead to improved guarantees for sketched algorithms in federated learning and bandits.
- GenAI Workshop on Advanced Techniques in Auto-prompt Tuning, Amazon Machine Learning Conference (AMLC)
TL;DR
Adapted the INSTINCT bandit-based prompt optimization framework for theft detection using VLMs, employing Mistral for prompt generation and Neural UCB for surrogate optimization, yielding a 3% AUC gain.
2024
- NeuRIPS, 2024
TL;DR
The high communication cost between the server and the clients is a significant bottleneck in scaling distributed learning for overparametrized deep models. One popular approach for reducing this communication overhead is randomized sketching. However, existing theoretical analyses for sketching-based distributed learning (sketch-DL) either incur a prohibitive dependence on the ambient dimension or need additional restrictive assumptions such as heavy-hitters. Nevertheless, despite existing pessimistic analyses, empirical evidence suggests that sketch-DL is competitive with its uncompressed counterpart, thus motivating a sharper analysis. In this work, we introduce a sharper ambient dimension-independent convergence analysis for sketch-DL using the second-order geometry specified by the loss Hessian. Our results imply ambient dimension-independent communication complexity for sketch-DL. We present empirical results both on the loss Hessian and overall accuracy of sketch-DL supporting our theoretical results. Taken together, our results provide theoretical justification for the observed empirical success of sketch-DL.
2022
- ALT, 2022
TL;DR
In this paper, we consider a bandit problem in which there are a number of groups each consisting of infinitely many arms. Whenever a new arm is requested from a given group, its mean reward is drawn from an unknown reservoir distribution (different for each group), and the uncertainty in the arm's mean reward can only be reduced via subsequent pulls of the arm. The goal is to identify the infinite-arm group whose reservoir distribution has the highest (1−α)-quantile (e.g., median if α=1/2), using as few total arm pulls as possible. We introduce a two-step algorithm that first requests a fixed number of arms from each group and then runs a finite-arm grouped max-quantile bandit algorithm. We characterize both the instance-dependent and worst-case regret, and provide a matching lower bound for the latter, while discussing various strengths, weaknesses, algorithmic improvements, and potential lower bounds associated with our instance-dependent upper bounds.