Research

1. Geometry in the token, not in the pipeline

   2026 · under review

   A VLM answers "is the chair in front of the sofa" from a grid of 2D patch features that carry no depth, so it answers from word statistics. We lift each patch to a metric 3D point and cluster the points into voxel tokens with a soft assignment. The question-answering loss trains the depth, the camera, and the grouping together.

   Reported

   Spatial reasoning improves with more spatial QA data, or with detection and segmentation estimators added at inference.

   Found

   Language loss alone learns a text shortcut: 41.5% on binary relations, below chance. Moving geometry into the tokenizer reaches 81.33% on SpatialRGPT-Bench at 1B parameters in 3.6 GB, with no region proposal at inference.

   Paper

2. What a benchmark score identifies

   2026 · under review

   MMVP's 6% for LLaVA-1.5 fits a model that cannot see. It fits equally a model that sees fine but always picks option (a). A single score is consistent with both readings until a control separates them. I build the cheapest control that does, then report what the published number was actually tracking.

   Reported

   LLaVA-1.5 scores 6.0% on MMVP, below the 25% random floor. The literature reads this as a perceptual deficit inherited from CLIP.

   Found

   93% of its failures pick the same slot whatever the image shows. On answers that survive reversal its perceptual accuracy is 70, indistinguishable from the model scoring 33.3.

   Paper

3. What a low-resource model is keying on

   Submitted to Transactions of the Association for Computational Linguistics, 2026

   Bangla encoders are evaluated on clean text. Bangla outside curated corpora is not clean. Rather than report that noise hurts, I control the comparison until the damage can be attributed to something, then test whether that attribution predicts which defense works.

   Reported

   Character noise degrades Bangla encoders because it fragments subword tokenization.

   Found

   With coverage held fixed, fragmentation predicts almost none of the loss (r = 0.05). The model is anchored to memorized entity spellings by 9.3 to 1 over context. Failing to find the entity at all accounts for 86% of the damage. Character augmentation recovers 14.3 F1, whereas gradient-based adversarial training (FGSM) recovers nothing outside seed noise.

   Paper

4. FaithAudit: Do explainable deepfake detectors cite the region that caused their verdict?

   2026 – present · PyTorch

   Explainable forgery detectors are scored for plausibility or localization overlap, never for whether the cited region drives the verdict. I test that with counterfactuals that contain no synthetic pixels: every face-swap frame has a frame-aligned authentic twin, so the authentic region spliced back in is the counterfactual. An identity splice controls for the seam. The codebase is complete and self-tested; first runs on FaceForensics++ are scheduled.

   CodeProject

5. PredictorProbe: Is the I-JEPA predictor spatially grounded?

   2026 – present · PyTorch

   JEPA models train a predictor to infer a hidden region's representation from context and a positional query, then discard it. Whether it uses the position it is asked about has not been quantified. I query the frozen predictor at a patch position and rank its output against every patch's true representation, which gives localization accuracy with no labels or training. The first real-weights run is complete.

   CodeProject

6. Adversarial Robustness in Low-Resource NLP (Bangla)

   2025 · PyTorch

   I ran character-, word-, and noise-level attacks on Bangla language models to measure which perturbations break which encoders. That work is the groundwork for the entity-orthography paper above.

   Code

7. Multiclass SQL Injection Detection with a 1D CNN–TinyBERT Model

   2024–2025 · PyTorch

   I built a 1D CNN–TinyBERT classifier for multiclass SQL-injection payloads. Because the classes are heavily imbalanced, I evaluated it with macro-F1 and PR-AUC under realistic adversarial scenarios.

   Project

8. Lightweight LLM and Reinforcement Learning for Text-to-SQL Generation

   2023 · PyTorch

   I built a TinyBERT text-to-SQL system on Spider. Reinforcement learning from execution feedback improved logical correctness without a larger model.

   Project

9. Undergraduate thesis: BCI-Oriented EEG Signal Analysis and Classification

   2022–2023 · TensorFlow

   I built end-to-end ERP detection on BCI Competition data, where cross-subject generalization was the bottleneck. The pipeline used FIR/notch/CAR preprocessing, PSD/WPD/STFT features, and NCA and Relief-F selection. The CNN reached 92% cross-subject accuracy. Ablations attributed most of the gain to preprocessing and feature selection rather than the classifier.

Publications

@article{sen2022bangla,
  author  = {Sen, Ovishake and Fuad, Mohtasim and Islam, Md. Nazrul and Rabbi, Jakaria and Masud, Mehedi and Hasan, Md. Kamrul and Awal, Md. Abdul and Fime, Awal Ahmed and Fuad, Md. Tahmid Hasan and Sikder, Delowar and Iftee, Md. Akil Raihan},
  title   = {Bangla Natural Language Processing: A Comprehensive Analysis of Classical, Machine Learning, and Deep Learning-Based Methods},
  journal = {IEEE Access},
  volume  = {10},
  pages   = {38999--39044},
  year    = {2022},
  doi     = {10.1109/ACCESS.2022.3165563}
}

@inproceedings{ahamed2022blood,
  author    = {Ahamed, Towsif and Islam, Md. Nazrul and Raju, S. M. Taslim Uddin and Hashem, M. M. A.},
  title     = {Blood Donor Arrival Forecasting Using Regression Model and Analysis of Donor Behavioural Pattern},
  booktitle = {2022 25th International Conference on Computer and Information Technology (ICCIT)},
  pages     = {897--902},
  year      = {2022},
  doi       = {10.1109/ICCIT57492.2022.10054710}
}

@article{fuad2021recent,
  author  = {Fuad, Md. Tahmid Hasan and Fime, Awal Ahmed and Sikder, Delowar and Iftee, Md. Akil Raihan and Rabbi, Jakaria and Al-Rakhami, Mabrook S. and Gumaei, Abdu and Sen, Ovishake and Fuad, Mohtasim and Islam, Md. Nazrul},
  title   = {Recent Advances in Deep Learning Techniques for Face Recognition},
  journal = {IEEE Access},
  volume  = {9},
  pages   = {99112--99142},
  year    = {2021},
  doi     = {10.1109/ACCESS.2021.3096136}
}

* Equal contribution.

Projects

Teaching

1. Lecturer

   Dept. of CSE, Southeast University, Dhaka, Bangladesh

   February 2024 – Present

   • Teaching: programming, numerical methods, and introductory AI
   • Supervise six undergraduates on Final Year Design Projects in machine learning and NLP, including PredictorProbe.

2. Lecturer

   Dept. of CSE, Uttara University, Dhaka, Bangladesh

   July 2023 – February 2024

3. Part-time Lecturer

   Dept. of CSE, Central University of Science and Technology, Dhaka, Bangladesh

   March 2023 – July 2023

Honors

1. Ranked 10th globally and 1st among Bangladeshi teams, Indian Rover Design Challenge (IRDC) 2020

   Mars Society South Asia — Team Durbar, KUET

   2020

2. Dean's Award

   Faculty of Electrical and Electronic Engineering, KUET

   AY 2018–2019 and 2019–2020

Service

1. Vice President

   Hardware Acceleration Club of KUET (HACK)

   March 2022 – March 2023

   • Ran the club's activities and events.
   • Led workshops on commonly used microcontrollers.

2. Programmer, Software Sub-Team

   KUET Mars Rover (Team Durbar)

   March 2019 – December 2022

   • Senior Programmer: rover navigation and computer vision, including a lightweight CNN for object recognition.
   • Junior Programmer: worked with Arduino and Raspberry Pi to implement key rover systems.

Contact

• nazrul.islam.cse.bd@gmail.com
• Google Scholar
• GitHub
• LinkedIn
• ORCID