Research
Research Experience
Research Intern — Dept. of Computer Science,
William & Mary
Under
Prof. Yanfu Zhang
| Nov 2025 – Present
- Developing a "Position-Aware" In-Context Learning (ICL) evaluation pipeline for long-context LLMs (Qwen2.5 7B) to mitigate performance degradation ("Context Rot") in multi-shot prompting.
- Engineering automated, large-scale inference pipelines via vLLM on NVIDIA RTX A6000 (48GB VRAM) to test whether ordering candidate examples using the "Lost in the Middle" attention bias yields zero-cost accuracy gains.
- Co-authored DyV2X, a dynamic graph learning framework for V2X cooperative trajectory prediction, submitted to IROS 2026 (2nd author).
- Designed a GRU history encoder across 6 CTDG architectures (TGN, TGAT, GraphMixer, DyGFormer, DyRep, JODIE) on the V2X-Seq-TFD dataset; achieved minADE=0.952m, minFDE=1.664m, MR=24.7%.
Publications
[Under Review]
DyV2X: Dynamic Interaction Graphs for V2X Cooperative
Trajectory Prediction
Yiran Ding, Muhammed Muminul Hoque,
Maiqi Jiang, Yongming Qin, Sidi Lu, Yanfu Zhang
IEEE/RSJ International Conference on Intelligent Robots and
Systems (IROS 2026)
[ECCE 2025]
Privacy-Preserving Digital Identity Verification:
AES-Encrypted Smart NID with Two-Factor Protection
Ekram Hossain, Puja Rani Saha,
Muhammed Muminul Hoque,
Md. Saiful Islam, Jonayed Al-Faruk, Md. Alamgir Hossain
4th International Conference on Electrical, Computer &
Communication Engineering |
[DOI]
[BIM 2025]
Toward Accurate Urban Temperature Forecasting in Seoul:
A Hybrid Ensemble with Explainable AI
Jonayed Al Faruk, Muhammed Muminul Hoque,
Md. Nawaz Shorif, Susmoy Paul, Md. Bodrul Islam, Joyonto Das
International Conference on Big Data, IoT and Machine Learning
[STI 2025]
Quantum-Resistant Blockchain with ROUND5
Lattice-Based Cryptography
Puja Rani Saha, Md. Saiful Islam, Tushar Kanti Saha,
Muhammed Muminul Hoque,
AKM Bahalul Haque, Ekram Hossain
7th International Conference on Sustainable Technologies
for Industry 5.0 |
[DOI]
[ECCT 2026]
A Comparative Analysis of Traditional and Ensemble
Machine Learning Models for Phishing Email Detection with
Explainable AI
Muhammed Muminul Hoque,
Md. Masudur Rahman, Jonayed Al-Faruk, AKM Bahalul Haque, Mir Tasrif Ahmed, MD. Nayeem Pervez
International Conference on Electrical, Computer and Communication Technologies (ECCT 2026)
Ongoing Research
[In Preparation]
ZEAL: A Zero-Shot Ensemble Framework for Adaptive Phishing Detection with LLMs and Structured Debate
(2025 – Present)
Muhammed Muminul Hoque
- Multi-tier detection pipeline combining adaptive keyword scoring, a 4-model LLM ensemble (GPT-4o-mini, Gemini, DeepSeek, LLaMA), and structured debate escalation — zero training data required.
- Achieves 98.44% accuracy and 97.46% F1-score on 4,812 emails across two benchmarks, competitive with fine-tuned RoBERTa baselines without any task-specific training.
- Dynamic keyword store learns emerging phishing vocabulary continuously via exponential time-decay weighting and accuracy-gated promotion.
.png)
.png)