|
"All models are wrong, but some are useful." – George Box [Name pronounced either as Saw-rv or Show-oo-rob (Bengali)] I am a Computer Science Ph.D. candidate at the University of Colorado Boulder, advised by Prof. Lijun Chen, and a Visiting Student Researcher in the Department of Computing and Mathematical Sciences at Caltech, working with Profs. Steven Low and Adam Wierman. My research focuses on the mathematical foundations of sequential decision-making under uncertainty, with a primary emphasis on Multi-Armed Bandits and Multi-Agent Reinforcement Learning (MARL). I investigate how structural, physical, and strategic constraints shape the limits of learnability, particularly in decentralized coordination, continuous action spaces, and evolving networked structures. More recently, I have also become interested in foundation models: the geometry of attention, including kernelized attention over symmetric cones, and the post-training and reinforcement learning methods that shape how these models reason and align. Prior to my Ph.D., I earned my Master’s in Computer Science at the University of Colorado Boulder and my bachelor's degree from the Birla Institute of Technology, Mesra. Before moving to the U.S., I worked as a Software Engineer at Flipkart in Bangalore, designing scalable search pipelines and centralized data recovery platforms. Beyond research, I find myself drawn to cinema, literature, and music. Cricket has been a constant since childhood: I played for a local club at school, later for the Flipkart corporate team, and now follow the game closely as a spectator. Fun Fact. My Erdős Number is 4 and Dijkstra Number is 5.
Email •
LinkedIn •
CV •
Google Scholar |
Taken in Ouray, Colorado, |
|
|
|
|
|
I develop the mathematical foundations of sequential decision-making to establish how structural, physical, and strategic constraints shape the limits of learnability in uncertain and stochastic environments. My overarching research agenda focuses on four core areas: (i) Scalable Multi-Agent Reinforcement Learning, where I develop theoretical frameworks to overcome the curse of dimensionality by exploiting policy-dependent locality; (ii) Graph-Constrained & Dynamic Bandits, characterizing the limits of online learning when an agent's actions are restricted by evolving network topologies; (iii) Continuous & Incentivized Learning, designing sequential decision-making algorithms for Lipschitz-structured action spaces and environments with biased feedback; and (iv) Foundation Models, where I study the geometry of attention, including kernelized linear attention over symmetric cones and the capacity limits it imposes, along with post-training and reinforcement learning methods for reasoning and alignment. * Equal contribution. |
| Publications & Preprints | |
|
Ayoub Ghriss, Sourav Chakraborty Under Review. arXiv (preprint)
We cast associative recall in linear attention as a spherical-packing problem and derive feature maps that certify nonnegative attention weights through self-dual symmetric cones. Rank-one PSD features enlarge the recurrent state without adding parameters, lifting linear attention's capacity limit.
|
|
Sourav Chakraborty*, Amit Kiran Rege*, Claire Monteleoni, Lijun Chen Accepted at IEEE Conference on Decision and Control (CDC) 2026 in Honolulu, Hawaii. arXiv / pdf
We study stochastic bandits on dynamic graphs where the learner is confined to local movement. A sliding-window mixing condition ensures stable exploration and navigation; local explore-then-commit algorithms provably achieve sublinear expected regret under this condition.
|
|
Sourav Chakraborty*, Amit Kiran Rege*, Claire Monteleoni, Lijun Chen Accepted at Learning for Dynamics & Control Conference (L4DC) 2026 in Los Angeles, California, United States. (Selected as an Oral Presentation.) ★ Best Paper Runner-Up (Top 1%) arXiv (preprint)
We formalize a novel bandit framework where arm availability is constrained by local moves on evolving random graphs (Erdős–Rényi and Edge-Markovian). We design a two-phase algorithm utilizing lazy random walks to achieve near-optimal, sublinear regret under these non-stationary physical constraints.
|
|
Sourav Chakraborty*, Amit Kiran Rege*, Claire Monteleoni, Lijun Chen Accepted at Learning for Dynamics & Control Conference (L4DC) 2026 in Los Angeles, California, United States. arXiv (preprint)
We establish that locality in decentralized MARL is inherently policy-dependent, introducing a novel spectral condition that decouples environmental coupling from policy smoothness. This framework enables policy improvement that circumvents the curse of dimensionality.
|
|
Sourav Chakraborty*, Amit Kiran Rege*, Claire Monteleoni, Lijun Chen Accepted at International Conference of Artificial Intelligence and Statistics (AISTATS) 2026 in Tangier, Morocco. arXiv (preprint)
We introduce a communication-free, modular protocol for decentralized multi-player bandits over continuous, Lipschitz-structured action spaces. Our approach decouples agent coordination from exploration, achieving near-optimal regret without horizon-dependent coordination costs.
|
|
Sourav Chakraborty*, Amit Kiran Rege*, Claire Monteleoni, Lijun Chen Accepted at ARLET Workshop at Neural Information Processing Systems (NeurIPS) 2025 in San Diego, California, United States. OpenReview
We study online learning where an agent is constrained to local movements on a dynamic graph sequence.
We formalizing sufficient structural conditions for learnability and design near-optimal, exploration policies.
|
|
Sourav Chakraborty*, Amit Kiran Rege*, Claire Monteleoni, Lijun Chen Accepted at IEEE Conference on Decision and Control (CDC) 2025 in Rio de Janeiro, Brazil. IEEE / arXiv
We develop incentivized exploration schemes in continuous metric spaces where the system compensates myopic agents to explore, despite receiving biased feedback. We develop near-optimal uniform discretization-based algorithms.
|
|
Sourav Chakraborty and Lijun Chen. Accepted at IEEE American Control Conference (ACC) 2024 in Toronto, Canada. Master's Thesis, Committee: Lijun Chen, Raf Frongillo, Bo Waggoner. IEEE version / arXiv / thesis
We address the challenge of incentivizing myopic agents to explore in both abruptly-changing and continuously-drifting bandit environments with drifted feedback and propose near-optimal algorithms.
|
|
Additional manuscripts currently under review are listed in my CV. |
|
|
|
Beyond my doctoral research at Boulder, I have worked in both academic research and industry settings. |
 |
Visiting Student Researcher at Caltech, Computing & Mathematical Sciences (2026)
|
 |
Software Engineer at Flipkart (2016-2019)
|
|
My teaching contributions span undergraduate computer science instruction and department-level instructional support, including course operations, student-facing instruction, assessment, staff training, and mentorship. |
 |
|
Last updated: November 2026
Thanks Jon Barron!