Md. Ismail Hossain

I am a researcher at the Machine Intelligence Lab, North South University, Bangladesh. I work on efficient post-training of language models, mainly knowledge distillation, on-policy self-distillation and sparsity. I am interested in where the learning signal actually matters when a model adapts, so that models can be improved with less data, annotation and compute.

Much of how I think about this comes from philosophy, psychology and cognitive science. People learn a great deal from a few well-chosen experiences and quietly ignore the rest. I want models to learn the same way.

Previously, I was a Research Associate at the Apurba-NSU R&D Lab, working with Dr. Nabeel Mohammed and Dr. Shafin Rahman, and a Fatima Fellow (Hugging Face & Google Colab) mentored by Dr. Isidora Tourni. I received my B.Sc. in Computer Science and Engineering from North South University in 2021.

I am open to PhD positions and research collaborations.

Email CV Google Scholar GitHub LinkedIn X

News

Sep 2026Released OASIS on arXiv, on why on-policy self-distillation fades at scale and how to fix it.
Jul 2026Presented LumiNet at ICML 2026.
Jun 2026Papers accepted to IEEE JBHI and Interspeech 2026.
May 2026Received a compute grant from Lambda.
Apr 2026Received a research grant from Adaption.
Sep 2025Two papers accepted to the EMNLP 2025 Industry Track.
Jul 2025LumiNet accepted at TMLR with J2C Certification.

Older news

Jul 2025Papers accepted at Scientific Data and IEEE SMC 2025.
Mar 2025Received project funding from the Research & Innovation Center (RIC), Bangladesh.
Jan 2025Paper accepted at IEEE Transactions on Artificial Intelligence.
Nov 2024Selected for the Oxford Machine Learning Summer School.
Sep 2024Started the Fatima Fellowship (Hugging Face & Google Colab).
Jul 2024Paper accepted at BMVC 2024.
Mar 2024Lightning talk on LumiNet at Cohere For AI; paper accepted at PML4LRS @ ICLR 2024.
May 2023Paper accepted at PLOS ONE.
Dec 2022Paper accepted at IEEE Access.
Sep 2022Second runner-up, DL Sprint ASR Contest 2022.
Jun 2022Paper accepted at ICPR 2022.

Selected Publications

* equal contribution. Full list on Google Scholar.

OASIS keeps its gain over the base model as model size grows, while OPSD's gain fades

Overcoming Scaling Limits in On-Policy Self-Distillation for LLM ReasoningNew Md. Ismail Hossain, Humaira Kousar, Isidora Chara Tourni arXiv preprint, 2026 paper

On-policy self-distillation stops helping as models grow, because most updates land on attempts that never reach a correct answer. OASIS distills only along verified attempts and keeps a gain of 3.2 to 3.8 points at every scale, using only final-answer labels.

LumiNet: Perception-Driven Knowledge Distillation via Statistical Logit Calibration Md. Ismail Hossain, MM Lutfe Elahi, Sameera Ramasinghe, Ali Cheraghian, Fuad Rahman, Nabeel Mohammed, Shafin Rahman TMLR 2025, J2C Certification; presented at ICML 2026 paper / code

Calibrates teacher and student logits through a shared statistical function, reducing the propagation of teacher errors in distillation. Received a TMLR J2C Certification.

COLT: Cyclic Overlapping Lottery Tickets for Faster Pruning of Convolutional Neural Networks Md. Ismail Hossain, Mohammad Rakib, MM Lutfe Elahi, Nabeel Mohammed, Shafin Rahman IEEE Transactions on Artificial Intelligence, 2025 paper / code

Uses the overlap between winning tickets of models trained on different tasks to prune about 50% faster than iterative lottery-ticket pruning.

3D Point Cloud Network Pruning: When Some Weights Do not Matter Amrijit Biswas*, Md. Ismail Hossain*, MM Lutfe Elahi, Ali Cheraghian, Fuad Rahman, Nabeel Mohammed, Shafin Rahman BMVC 2024 paper / code

About 99% of the weights in 3D point-cloud networks can be removed with near-baseline accuracy.

CAPSTONE: Composable Attribute-Prompted Scene Translation for Zero-Shot Vision–Language Reasoning Md. Ismail Hossain*, Shahriyar Zaman Ridoy*, Moshiur Farazi, Nabeel Mohammed, Shafin Rahman EMNLP 2025, Industry Track paper / code

Replaces a VLM's vision encoder with a composition of small vision models whose outputs an LLM reasons over.

Uncovering Critical Features for Deepfake Detection through the Lottery Ticket Hypothesis Lisan Al Amin*, Md. Ismail Hossain*, Thanh Thi Nguyen, Tasnim Jahan, Mahbubul Islam, Faisal Quader IEEE SMC 2025 paper

Mitigating Carbon Footprint of Hyper-parameter Selection During Knowledge Distillation Kazi Rafat, Sadia Islam, Abdullah Al Mahfug, Md. Ismail Hossain, Fuad Rahman, Sifat Momen, Shafin Rahman, Nabeel Mohammed PLOS ONE, 2023 paper / code

Other Publications

A Point Cloud Transformer for Remote Monitoring and Automated Assessment of Physical Rehabilitation Exercises. K. Rafat, M. I. Hossain, M. M. Elahi, S. Momen, F. Rahman, N. Mohammed, S. Rahman. IEEE JBHI, 2026.
Dynamic Block-Online Streaming ASR for Low-Resource Agglutinative Code-Switching Speech with Morphology-Aware Evaluation. K. Rafat, A. Imran, M. I. Hossain, M. R. Ali, S. Rahman, N. Mohammed. Interspeech 2026.
AutoDSPy: Automating Modular Prompt Design with Reinforcement Learning for Small and Large Language Models. N. Azim, A. U. Alam, H. B. Omar, A. M. M. A. Jami, J. I. Ahad, M. R. Kabir, M. I. Hossain, M. R. Amin, S. Rahman, N. Mohammed. EMNLP 2025, Industry Track.
A Multidimensional Dataset of Student Gaming Behavior and Academic Correlates in Bangladesh. M. I. Hossain, L. A. Amin, M. U. Emon, P. V. Chauhan, A. Rudrabatla, F. Quader. Scientific Data, 2025.
Precision-Driven Low-Resource Speech Synthesis for Bangla Text-To-Speech System. T. S. Shahjahan, M. I. Hossain, K. Rafat, M. R. Amin, F. Rahman, N. Mohammed. PML4LRS Workshop @ ICLR 2024.
LILA-BOTI: Leveraging Isolated Letter Accumulations By Ordering Teacher Insights for Bangla Handwriting Recognition. M. I. Hossain, M. Rakib, S. Mollah, F. Rahman, N. Mohammed. ICPR 2022.

Experience

Machine Intelligence Lab, North South University. Researcher.Present
Fatima Fellowship, Hugging Face & Google Colab. Research Fellow, mentored by Dr. Isidora Tourni.2024 – 2025
Apurba-NSU R&D Lab, North South University. Research Associate.2021 – 2025

Awards & Grants

Research Grant, Adaption2026
Compute Grant, Lambda2026
J2C Certification for LumiNet, TMLR2025
Research Innovation Center Fund, World Bank / RIC Bangladesh2025
Fatima Fellowship, Hugging Face & Google Colab2024
Oxford Machine Learning Summer School, selected participant2024
DL Sprint ASR Contest, second runner-up2022

Service & Talks

Reviewer: NeurIPS 2026, CVPR 2026, ICLR 2026 and 2027, TMLR, NeurIPS 2025, ICML 2025 workshop (LXAI), NeurIPS 2024 workshop (ML & Compression).

Talks: LumiNet, Cohere For AI (2024); Efficiency in deep learning, NSU ACM Student Chapter (2022); ASR and OCR, Microsoft Learn Student Ambassadors (2022).

Mentoring: I have mentored undergraduate researchers at the Apurba-NSU R&D Lab, and I am happy to talk with first-generation students starting in research.