Provost's Postdoctoral Fellowship
2026 – 2028
College of Engineering, University of Notre Dame
Funding
NAVSEA & STTR awards
Ph.D., GMU
Naval Sea Systems Command; multiple STTR awards
I enjoy the proposal side of research as much as the papers, and I am always building toward the next one.
If you are interested in applying for grants as co-PIs, please knock me: mraihan@nd.edu.
Thread one
Code LLMs & program synthesis
How code models behave, and how to make them better, outside English and outside Python. This thread spans dedicated model families, instruction corpora, and execution-based evaluation for low-resource settings.
LREC 2026
The first dedicated family of Code LLMs for Bangla (1B and 9B), built on a 300K instruction–code corpus, with 11–18% Pass@1 gains over larger baselines on MBPP-Bangla.
NAACL 2025
Extends HumanEval to 204 natural languages and 25 programming languages, 836,400 prompts in total, so code generation can be measured far beyond English and Python.
NAACL 2025
The first language model and evaluation suite for the Mojo programming language, a case study in adapting Code LLMs to a brand-new language.
BigData 2024
A structured map of the Code LLM landscape: architectures, training data, and evaluation practices.
BLP 2025
The first shared task on Bangla code generation, which I organized, with 158 international participants.
Thread two
Multilingual & low-resource NLP
Modeling and evaluation across natural languages, with a long-running focus on Bangla and on code-mixed text, plus how multilingual LLMs hold up in sensitive domains.
ACL 2025
Open Bangla LLMs, pretrained and instruction-tuned, that outperform prior open alternatives for the language.
EACL 2026
Evaluates proprietary and open LLMs on eight mental health datasets across languages and their machine-translated counterparts (LoResLM workshop).
BLP 2023
Introduces TB-OLID and shows how offensive language behaves once Bangla is romanized and mixed with English (BLP @ EMNLP 2023).
2023 – 24
Trilingual Bangla-English-Hindi code-mixed corpora for sentiment, offensive language, and emotion detection (AACL 2023; LREC-COLING 2024).
Thread three
LLM safety & AI in CS education
Guardrails for code assistants and the role of LLMs in introductory computing. This is the thread my Notre Dame fellowship extends, toward safety-by-construction guardrails for Code LLMs.
EACL 2026
A prompt taxonomy, an 8K-prompt dataset, and PromptShield (0.93 F1), which together cut harmful code completions by 30–65% while preserving legitimate coding help (Findings of EACL 2026).
SIGCSE 2025
A systematic review of how LLMs are actually being used, and misused, in CS classrooms (SIGCSE-TS 2025).
JIIS 2025
Measures how far current LLMs get on real introductory assignments, and where they quietly fail (Journal of Intelligent Information Systems, Springer).
SemEval 2024
Transformer-based detection of machine-generated text, an early ingredient of the integrity side of this thread.
Thread four
Benchmarks & datasets
Large, openly released resources that make under-tested settings measurable. Most of my modeling work starts by building the yardstick first.
NAACL 2025
836,400 prompts across 204 natural and 25 programming languages, the widest execution-based code generation benchmark of its kind.
ISMIS 2024
A benchmark of real introductory CS prompts for testing how LLMs handle classroom-style programming questions.
LREC 2024
A large multi-task dataset for mental health signals in social media text (LREC-COLING 2024).
BEA 2024
An extensible, massively multilingual lexical simplification pipeline dataset and the BEA 2024 shared task built on it.