Statement & publications
Research directions
Statement of research
I study when large models deserve to be trusted. I build benchmarks for high-stakes settings, from medical visual questions to prescription errors, that measure what language and vision-language models actually catch. Now I am asking whether models can keep improving after deployment, through test-time reinforcement learning with verifiable rewards.
Read the full research statement (PDF) ↗
Interests
Evaluation and benchmarking of LLMs and vision-language models in high-stakes domains; test-time and continual reinforcement learning for LLMs with verifiable rewards; and parameter-efficient fine-tuning (LoRA/PEFT).
Also: low-resource and Bangla NLP, medical visual question answering, retrieval-augmented generation, and robustness of federated learning.
Thread 01
High-stakes evaluation
Can a vision-language model read a prescription or a clinical image and catch what matters? Usually not as well as its benchmarks suggest.
Thread 02
Test-time reinforcement learning
Can a model keep improving after deployment from rewards it can verify, across languages, domains, and modalities?
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.
“Where does a model fail when the stakes are real, and can it learn to stop failing without a human in the loop?”
Publications
Published
2026 · Conference · Findings of the Association for Computational LinguisticsPublished
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.
2026 · Workshop · Second AAAI Bridge Program on AI for Medicine and HealthcarePublished
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.
2025 · Workshop · 2nd BLP (BLP-2025), IJCNLP-AACLPublished
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.
2025 · Workshop · 2nd BLP (BLP-2025), IJCNLP-AACLPublished
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.
Under review
2026 · Conference · ACL Rolling Review (October 2026), targeting NAACL 2027Under review
RxFact: Can Large Language Models Detect Clinically Harmful Errors in Prescription Images?
Intesar Tahmid · Akm Moshiur Rahman Mazumder · Mir Sazzat Hossain · Md. Sadman Sakib · S.M. Shahriar · Faria Iffat Reshma · Md Fahim · Ashraful Islam · Md Farhad Alam Bhuiyan
A benchmark of prescription images testing whether large vision-language models can detect clinically harmful errors before they reach a patient.
pdf on request
2026 · Journal · Computer Networks (Elsevier)Under review
AD-SFL: Activation-Space Anomaly Detection and Reinforcement Learning-Based Defense Against Data Poisoning in SplitFed Learning
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.
pdf on request
In preparation
Drafts in progress. Details available on request.
StreamTTRL: Continual Test-Time Reinforcement Learning Across Languages, Domains, and Modalities
Intesar Tahmid et al. · Targeting ACL 2027
❦ ❦ ❦
Service
- Present and critically evaluate ML and network-security papers in the Green Networking Research Group reading sessions
- Collaborate on ideation, experimental design, and peer review of ongoing work within the group
- Co-organized seminar series, CSEDU