01
Understand
the prediction.
Token attribution, mechanistic interpretability, and formal checks of reasoning. Which parts of a model’s input and computation matter?
Explore interpretability HETA · NSF-CoTPh.D. student · University of Florida
Vishal
Pramanik.
Understanding AI.
Building trust.
I study what happens inside AI models—and how to make their decisions and collaborations more reliable.
Advised by Prof. Sumit Kumar Jha
Curious about the
mechanisms behind the answer.
mechanisms behind the answer.
Looking for a Summer 2027 research internship in LLM/VLM interpretability, trustworthy AI, or multi-agent collaboration.
Let’s connect01 / A little about me
I ask AI models
“But why?”
A convincing answer is a starting point.
I want to understand the mechanism behind it.
I’m a Ph.D. student in Computer and Information Sciences at the University of Florida, working with Prof. Sumit Kumar Jha. My research connects explanations of model behavior with interventions: from token attribution and machine unlearning to jailbreak defense and reliable agent coordination.
Before UF, I studied at Florida International University, earned my M.Tech. at IIT Bombay, and spent two years developing server firmware at Intel. That path—from hardware to model internals—shapes how I think about dependable AI.
02 / Research directions
From explanations
to interventions.
Three connected questions guide my work, across language models, vision models, and teams of agents.
Understand
02
Intervene
with intention.
Representation steering, jailbreak analysis, and machine unlearning. How can we change a model’s behavior precisely and responsibly?
Explore model safety Nullspace Steering · Selective Amnesia03
Keep agents
working together.
Distributed intent, decentralized coordination, and long-horizon control. How do we keep a team of AI agents safe, on task, and in role?
Explore multi-agent systems CITADEL · URCP · ATLAS · ATC03 / Selected publications
Ideas, put to the test.
8 selected papers
CITADEL
Harmless in Pieces, Harmful in Motion: Detecting Multi-Agent Jailbreaks
Detecting harmful intent that emerges across agents, tools, and shared memory—even when the individual steps appear harmless.
URCP
Unified Resource-Grounded Coordination Protocol for Orchestrator-Free Heterogeneous Multi-Agent Systems
How can heterogeneous agents agree on roles, share resources, and repair failures without a central orchestrator?
ATLAS
Drift No More: Spectral Locking for Long-Horizon Open-Weight Multi-Agent AI
Intervening on hidden representations to keep collaborating agents aligned with their task, role, and shared state.
Selective Amnesia
Selective Amnesia using Contrastive Subnet Erasure for Class Level Unlearning in Vision Models
Making vision models forget a selected class through targeted subnetwork modification.
HETA
Hessian-Enhanced Token Attribution (HETA): Interpreting Autoregressive LLMs
Understanding which input tokens influence a language model’s predictions through second-order attribution.
Nullspace Steering
Jailbreaking the Matrix: Nullspace Steering for Controlled Model Subversion
Studying how interventions in internal representation spaces expose vulnerabilities in language-model safety.
NSF-CoT
NSF-CoT: Neuro-Symbolic Formal Verification of Chain-of-Thought Faithfulness in Contextual Question Answering
Combining neuro-symbolic reasoning and formal checks to examine whether explanations support an answer.
ATC
ATC: Agent Trajectory Control for Role-Contract Enforcement in Black-Box Multi-Agent Systems
Enforcing agent role contracts through output scoring and constrained regeneration when model internals are inaccessible.
Explore each paper’s research page, with full PDFs and code linked where available.
Explore the work
From the paper to the project.
04 / The path so far
A researcher
with an
engineering lens.
From server memory to model memory, I’m drawn to understanding how complex systems work.
Read my CV / resume- 2026 — Present
University of Florida
Ph.D. student · Graduate Research Assistant
Computer and Information Sciences
Advisor: Prof. Sumit Kumar Jha - 2024 — 2025
Florida International University
Doctoral study · Transferred to UF
Computer and Information Sciences · GPA 4.0/4.0
- 2022 — 2024
Intel
Silicon Firmware Development Engineer
DDR5 memory initialization, server boot flows, and performance optimization.
- 2020 — 2022
IIT Bombay
M.Tech. · Computer Science & Engineering
GPA 9.01/10 · Certificate of Excellence in Research
- 2015 — 2019
West Bengal University of Technology
B.Tech. · Computer Science & Engineering
GPA 8.9/10
Recognition & service
Along the way.
2026CVPR HighlightClass-level machine unlearning
IIT BombayExcellence in ResearchMaster’s thesis research
2024Second Place Poster AwardIEEE NAECON
ServiceResearch peer reviewICLR · NeurIPS · AAAI · CVPR
05 / Let’s connect
Better questions.
Better AI.
I’d love to work with people asking hard questions about how AI works—and how we can make it more trustworthy.
vishalpramanik@ufl.edu