Maximizing happiness
Investigating how students think, where they get stuck, and whether AI feedback actually helps.
I combine cognitive diagnostic modeling with data from students’ programming attempts to study how AI feedback can improve understanding.
Findings of EMNLP, ACM Transactions, TMLR, and workshops at ICSE, ASE, and EDM
My work spans the classroom and frontier AI development. At the University of Dhaka, where I am an Assistant Professor of Computer Science & Engineering, I teach programming, build learning platforms, and supervise undergraduate thesis research. All of my recent research has grown out of that supervision. I also graduated at the top of my class in both my B.Sc. and M.Sc. here.
In industry, I spent three years at Turing as a program delivery manager for a frontier AI lab, working across robotics, computer-use agents, and reasoning models. I now consult for Pareto on software engineering training data. Across these settings, my research asks how AI changes the skills students need and how we can use it to deepen their understanding.
“Maximizing happiness” is the idea I try to put into practice through teaching and widening access to education. I co-founded AlterYouth, which funds 2,700+ monthly scholarships for underprivileged primary school students in Bangladesh.
I collaborate with researchers across universities in Bangladesh and abroad, in low-resource language technology, educational data mining, software engineering, and language-model evaluation.
I am open to research collaboration in any of these areas.
How can we infer students’ programming skills from their work, and evaluate whether AI feedback addresses their misconceptions?
Current research on skill diagnosis, learning analytics, and tutor feedback. These projects investigate how to make AI-supported assessment useful for teaching.
I compare two cognitive diagnostic models, DINA and G-DINA, against three small and two large language models on 93 students, 49 problems, and a Q-matrix mapping problems to 12 skills. The aim is to test whether smaller, more accessible models can support fine-grained skill diagnosis.
I reconstruct pre-acceptance trajectories from 6,491 submissions by 852 students across 60 problems. These histories provide evidence of how students revise their solutions and where difficulties persist across attempts.
I evaluate whether AI programming tutors detect, repair, or reinforce the misconception behind an incorrect solution. I score 2,350 tutor responses drawn from 3,101 annotated submissions, across three models and four prompting strategies.
Datasets and evaluation for language technology that can interpret Bangla in its cultural context.
We annotated 3,000 Bangla memes for metaphor, cultural reference, and intent. The dataset supports evaluation of how multimodal models interpret the relationship between images and language. The paper has been accepted to Findings of EMNLP 2026.
I mainly teach programming courses at CSEDU, from introductory programming and algorithms to software design and web and app development. I build platforms for students to practise, receive feedback, and learn with AI.
I served as a co-coach for the University of Dhaka at the 46th ICPC World Finals in Luxor, Egypt.
Website & Registration ChairCurrently serving at 3rd IEEE COMPAS 2026
Member SecretaryVisionX 2025
Problem Setter and Judge CoordinatorCode Samurai
Member Secretary, Judge Panel and Problem Setter CommitteeICPC Dhaka Regional
First Runner Up · Health CategoryEATL–Prothom Alo Apps Contest 2016
Champion · Best Student SoftwareRed Hat Code Healthy with OpenShift
Best Mobile ApplicationMicrosoft VS Marketplace Hackfest · Java Tools Challenge
Selected papers on programming assessment, student submission data, and language-model evaluation, alongside work on Bangla language understanding.
Selected papers below, followed by the complete list, including accepted work.
Strict Graders and Hallucinated Mastery: A Cognitive Diagnostic Study of Language Models on Dynamic Programming
CSEDM at EDM 2026 · To appear
Language-model graders swing between marking too hard and crediting skills the submission never shows.
Data in Brief
Provides sequential verdict traces and attempt histories for studying how students revise programming solutions.
A-ProS: Towards Reliable Autonomous Programming Through Multi-Model Feedback
ACM Transactions on Software Engineering and Methodology
Uses cross-model feedback to improve the reliability of autonomous code generation.
POVC at ASE 2026 · To appear
Evaluates AI-generated unit tests across 15 student-developed repositories, examining assertion strength, edge cases, and maintainability alongside execution success.
Findings of EMNLP 2026 · To appear
A 3,000-item explainable dataset for culturally grounded interpretation of Bangla memes. Accepted to Findings of EMNLP 2026.
UHGMiner: A Framework and Algorithm for High-Utility Hypergraph MiningADMA 2026 · Jawad, Habib, Alam, Arefin, Ahmed, Leung
BanglaMemeX: Advancing Cultural Metaphoric Image Interpretation in Bangla with a Multimodal Explainable DatasetFindings of EMNLP 2026 · Sakib, Mahmud, Arefin, Fahim
VibeCheck: Assessing the Quality of LLM-Generated Unit Tests: A Multi-agent Empirical Study across Heterogeneous Repositories · arXivPOVC at ASE 2026 · Tabassum, Intesum, Arefin, Zaman
Strict Graders and Hallucinated MasteryCSEDM at EDM 2026 · Solaiman, Arefin
CHARD: Characteristic-Aware Hierarchical Detection for Bangladesh Road ScenesIEEE Access · Hasan, Fahad, Arefin, Khan
Hyperedge Anomaly Detection with Hypergraph Neural NetworkTransactions on Machine Learning Research · Alam, Rahman, Arefin, Ahmed, Mahmud, Sakib, Leung
A-ProS: Towards Reliable Autonomous Programming Through Multi-Model Feedback · arXivACM TOSEM · Tabassum, Hossain, Arefin, Islam, Zaman
CodeStream: A Dataset of Iterative Programming Submissions with Sequential Verdict Traces and Attempt HistoriesData in Brief · Lina, Mahmud, Hasan, Arefin, Rizvee, Rahman, Khan
LLM-ProS: Analyzing Large Language Models’ Performance in Competitive Problem SolvingLLM4Code at ICSE 2025 · Hossain, Tabassum, Arefin, Zaman
A New Tree-Based Approach to Mine Sequential PatternsExpert Systems with Applications · Rizvee, Ahmed, Arefin, Leung
Unveiling Market Sentiments: Stock-Market Responses to Diverse News EventsIMCOM 2024 · Solaiman, Rahman, Arefin, Ahmed, Leung, Madill
Parallelized 0/1 Knapsack Algorithm Optimization in CPU-GPU-Based SystemsIMCOM 2024 · Abid, Shaha, Kabir, Arefin
Applying Sequential Correlation to Identify Order Dependent Activity PairsDhaka University Journal of Applied Science and Engineering · Arefin, Islam, Ahmed
Mining Contextual Item Similarity without Concept HierarchyIMCOM 2022 · Arefin, Ahmed, Rizvee, Leung, Cao
A Robust Objective-Focused Algorithm to Detect Source-Code PlagiarismIEEE UEMCON 2022 · Rizvee, Arefin, Abid
Comparison of Different Sentiment-Analysis Techniques for Bangla ReviewsIEEE R10 HTC 2022 · Jabin, Suhi, Arefin, Hasib
An Automated System to Calculate Marks from Answer ScriptsIEEE IEMCON 2022 · Rizvee, Arefin, Khan, Islam, Rabbi
Comparative Analysis of Different Text Summarization Techniques Using Enhanced TokenizationSTI 2021 · Islam, Hossain, Arefin
Data-Driven Hybrid Optimisation-Based Deep Network for Short-Term Residential Load ForecastingIEEE IEMCON 2021 · Sakib, Hasib, Tasawar, Tanzeem, Arefin, Islam
Tree-Miner: Mining Sequential Patterns from SP-TreePAKDD 2020 · Rizvee, Arefin, Ahmed
Discovering Interesting Knowledge from Transactional DatabaseBangladesh Computer Society Journal of Computer and Information Technology · Vol. 2, pp. 33–40 · Arefin, Ahmed
Mining Sequential Correlation with a New MeasureICDM 2018 · Arefin, Islam, Ahmed
I am interested in research collaborations on student modeling, the changing skills needed in programming education, and how AI tutors can support students’ understanding. If these questions connect with your work, I would be glad to discuss them.
Attending EMNLP 2026 in Budapest. If you’ll be there, I’d be glad to meet and talk research.