Controller aware language model
How a language model reasons depends on the system guiding it. We train models to cooperate with reasoning controllers and generalize to unfamiliar inference workflows.
PhD candidate · Computer Science · UMass Amherst
Moumita Choudhury
I study how AI agents work, how they work together, and how they affect others.
At UMass Amherst’s Resource Bounded Reasoning Lab, I work with Prof. Shlomo Zilberstein. I completed my M.Sc. in Computer Science at UMass in 2024 and was a research intern with the Trustworthy AI group at Mitsubishi Electric Research Laboratories in 2025.
Before UMass, I was a research assistant at the University of Dhaka’s Cognitive Agents and Interaction Lab (CAIL) and a junior lecturer at Ahsanullah University of Science and Technology.
Outside the lab I enjoy singing, painting, and travelling.
Questions that keep me curious
Intelligence is a shared endeavor.
How do agents work?
How do they work together?
Distributed Constraint Optimization
Each agent makes a local decision, but the choices must work together. My DCOP research develops distributed algorithms for coordinating these decisions.
The Benefits and Limits of Trust Models
Why learning whom to trust does not always lead to successful agreements in repeated LLM negotiation.
How do they affect others?
Minimizing Negative Side Effects
Combining human feedback with distributed coordination to reduce joint side effects while limiting losses in task performance.
The Helpful Agent Problem
How agents exchanging favors can bypass intended workflows and harm others, even while following their local permissions.
Journal articles, conference papers, workshops, and preprints.