Statement & publications
Statement of research
I study when large models deserve to be trusted, and what it takes to make them dependable. My work measures what multilingual and multimodal systems actually understand rather than what their benchmark scores imply, follows how that understanding shifts as models are adapted to new domains, and asks whether the resulting gains survive distribution shift and adversarial pressure. I am drawn to problems where a careful diagnosis outperforms a larger model: a probe that reveals what fine-tuning obscures, or a defence that costs less than the attack it absorbs.
Thread 01
Visual question answering
Can a vision-language model read a clinical image and answer a question about it? Usually not as well as its benchmarks suggest.
Thread 02
LLM behaviour across stages
What a model knows with a frozen backbone, what changes under LoRA, and what only full fine-tuning reaches.
Thread 03
Parameter-efficient adaptation
Staging linear probing, fine-tuning, and LoRA so each does the part it is actually good at.
Thread 04
Network security with RL
Detecting poisoned clients in split and federated training, and learning how aggressively to exclude them.
“What does a model actually know at each stage of adaptation, and what is the cheapest intervention that closes the gap?”
Rafid Ahmed* · Intesar Tahmid* · Mir Sazzat Hossain · Tasnimul Hossain Tomal · Md Mahir Jawad · Anam Borhan Uddin · Md Fahim · Md Farhad Alam Bhuiyan (*equal contribution)
A rigorous evaluation of large vision-language models on Bangla medical VQA, exposing systematic failures in clinical terminology translation, visual grounding, and medical reasoning.
Tasnimul Hossain Tomal* · Intesar Tahmid* · Palash Roy · Mehedi Hasan · Md Abdur Razzaque (*equal contribution)
Poisoned clients in SplitFed Learning betray themselves in activation space. We detect them there, then let a reinforcement-learning policy decide how aggressively to exclude them, defending accuracy without assuming how many attackers there are.
Rafid Ahmed* · Intesar Tahmid* · Mir Sazzat Hossain · Tasnimul Hossain Tomal · Md Fahim · Md Farhad Alam Bhuiyan (*equal contribution)
Benchmarks LLM performance on Bangla medical visual questions, contributing a new clinically validated dataset. Reveals that frontier models struggle on clinical terminology and visual grounding.
Tasnimul Hossain Tomal* · Anam Borhan Uddin* · Intesar Tahmid · Mir Sazzat Hossain · Md Fahim · Md Farhad Alam Bhuiyan (*equal contribution)
A three-stage PEFT framework in which a linear probe initializes, full fine-tuning adapts, and LoRA efficiently specializes, yielding consistent gains on Bangla NLP tasks with far fewer trainable parameters.
Intesar Tahmid* · Rafid Ahmed* · Md Mahir Jawad · Anam Borhan Uddin · Md Fahim · Md Farhad Alam Bhuiyan (*equal contribution)
Submission to BLP-2025 Shared Task 1. Demonstrates that frozen-encoder probes are a strong, sample-efficient baseline for Bangla hate-speech classification.
❦ ❦ ❦
Currently reading
Service
A full-length research statement is available on request.
© 2026 Intesar Tahmid