Vishal Pramanik
PhD. Student @ UF · Location - Florida, US
I Ask AI Models 'But Why?' So You Don't Have To
About
I'm Vishal Pramanik, a PhD student in Computer and Information Sciences at the University of Florida, where I conduct research at the intersection of AI safety, explainable AI, and machine learning security under the guidance of Prof. Dr. Sumit Kumar Jha. My work focuses on understanding the inner mechanisms of deep learning models—from vision transformers to large language models—using mechanistic interpretability, attribution methods, and circuit-based approaches.
My research spans multiple critical areas in AI safety: developing novel attribution techniques for model interpretability, exploring adversarial robustness and jailbreak vulnerabilities, advancing machine unlearning frameworks, and investigating hyperdimensional computing for efficient representation learning. I'm particularly interested in moving beyond surface-level model analysis to understand the fundamental mechanisms that drive neural network behavior and decision-making.
Before joining UF, I earned my Master's degree in Computer Science Engineering from the Indian Institute of Technology Bombay, where I received the Certificate of Excellence in Research for my thesis work on natural language processing using large language models. I also spent two years at Intel Bangalore as a Silicon Firmware Development Engineer, contributing to memory initialization and optimization for DDR5-based server systems.
Education
Doctor of Philosophy (PhD) in Computer and Information Sciences
University of Florida
2025 - Present
Master of Technology (M.Tech.) in Computer Science Engineering
Indian Institute of Technology Bombay
2020 - 2022
GPA: 9.01/10.0
Bachelor of Technology (B.Tech.) in Computer Science Engineering
West Bengal University of Technology
2015 - 2019
GPA: 8.9/10.0
Research Experience
PhD Graduate Research Assistant
University of Florida
2025 - Present
- Conduct research on explainability, security, and unlearning in deep learning models
- Develop novel attribution methods to enhance interpretability of LLMs and vision models
- Focus on mechanistic and circuit-based approaches to understand inner workings of neural networks
- Explore applications spanning adversarial robustness, jailbreak analysis, and machine unlearning in both vision and language domains
- Investigate hyperdimensional computing frameworks for efficient representation learning and unlearning
- Leverage attribution techniques to provide deeper insights into model behavior and decision-making processes
Master's Research Assistant
Indian Institute of Technology Bombay
2020 - 2022
- Conducted research in Natural Language Processing using transformer-based models
- Developed novel techniques for language generation, sentiment analysis, information extraction, and machine translation
- Developed end-to-end movie script generation system for ErosNow, enabling users to generate scripts from storylines
- Developed novel AEG model using collaborative and transfer learning, achieving 85.50% accuracy and outperforming state-of-the-art approaches
- Published research paper "Hey AI, Can You Grade My Essay?: Automatic Essay Grading" in ICAAAIML 2025 - Springer Proceeding
Technical Experience
Silicon Firmware Development Engineer
Intel Bangalore, India
July 2022 - July 2024
- Contributed to server development team, specializing in main memory initialization during server booting process
- Developed initialization processes for unregistered DIMMs, including generating reference voltage tables for memory training
- Contributed to Memory Reference Code (MRC) development for servers with DDR5 DIMMs, focusing on boot flow, error handling, memory mapping, temperature control, and performance optimization
- Worked with validation team to design and conduct stress tests on various delay parameters of server processors
- Optimized server overclocking by adapting technology from standalone workstations, enhancing system throughput and performance
Publications
Hessian-Enhanced Token Attribution (HETA): Interpreting Autoregressive Language Models
NeurIPS 2025 Workshop: First Workshop on Foundations of Reasoning in Language Models
Jailbreaking the Security of LLMs: How Targeted Nullspace Perturbations Unmask Hidden Vulnerabilities
AAAI-26 Workshop on Artificial Intelligence for Cyber Security (AICS), 2026
SPUN: Spectral Projection–based UNlearning in Hyperdimensional Computing
AAAI-26 Workshop on Artificial Intelligence for Cyber Security (AICS), 2026
Fact or Hallucination? An Entropy-Based Framework for Attention-Wise Usable Information in LLMs
AAAI-26 Workshop on Artificial Intelligence for Cyber Security (AICS), 2026
Kurosawa: A Script Writer's Assistant
ACL Anthology | PDF
Can Linear Attributions Explain Nonlinear LLMs?
AAAI 2026 Workshop on Trust and Control in Agentic AI
Human Fatality Estimation in Aircraft Accidents
IEEE National Aerospace and Electronics Conference (NAECON), 2024 | PDF
Skills
Main technologies I work with:
Awards & Achievements
- Certificate of Excellence in Research - Computer Science Department, Indian Institute of Technology Bombay, for master's thesis on Natural Language Processing using Large Language Models
- Second Place Award - IEEE National Aerospace and Electronics Conference (NAECON) 2024, for poster presentation on "Human Fatality Estimation in Aircraft Accidents"
Project Webpages
Explore interactive webpages for my recent research work: