MD Azizul Hakim NeurIPS 2026 Sole Author
Portrait of MD Azizul Hakim

MD Azizul Hakim

NeurIPS 2026 Sole Author

AI ResearcherReinforcement Learning · Looped Language Models · Training & Inference Efficiency

I am an AI researcher working on reasoning efficiency in large language models: how the way a model is trained decides how well it spends compute at inference time. I am the sole author of a NeurIPS 2026 main-track paper, which shows across 50 models that training quality, more than parameter count or token budget, sets when extra reasoning stops paying off. My current work follows three threads:

  1. Training quality and test-time reasoning. When do extra reasoning tokens help, waste compute, or hurt accuracy, and which training choices move that boundary.
  2. Reinforcement learning for reasoning. Reward signals for RL post-training that teach reasoning models to stop once they have the answer, instead of over-generating.
  3. Looped language models. Models that reuse the same layers several times: whether a model should think deeper or longer, and how it can choose its own number of loops per input.

Earlier, I built bio-inspired, self-repairing neural networks and studied how intrusion detectors fail across industrial networks, which led to two first-author papers in Scientific Reports (Nature Portfolio) and three at IEEE QPAIN 2025. I am also a Kaggle Notebooks Master.

I hold a Diploma-in-Engineering in Computer Science & Technology from Bangladesh Sweden Polytechnic Institute, and I am looking for a fully funded BSc in Computer Science to continue this research. If you work on these questions, I would be glad to hear from you at azizulhakim8291 at gmail dot com.

News

Earlier news

  • Completed my Diploma in Computer Science & Technology at Bangladesh Sweden Polytechnic Institute.

  • Multi-Scale Temporal Homeostasis Enables Efficient and Robust Neural Networks released on arXiv.

Publications

1
NeurIPSsole author
2
Scientific ReportsNature Portfolio
3
IEEE QPAINconference papers
2
arXivpreprints

First author on every paper. Full list on Google Scholar.

  1. NeurIPS 2026

    Training Quality Determines Efficiency Boundaries in Test-Time Reasoning

    MD Azizul Hakim (sole author)

    Across 50 language models (0.5B–685B) and 67,000+ evaluations, training methodology rather than size or token budget decides where extra reasoning tokens stop helping. Capping generation length lifts accuracy by 15.4 points.

    acc. short ~1k tokens long budget higher training quality lower training quality
    Schematic of the finding: accuracy stops improving near 1,000 generated tokens, and training quality decides how early and how high a model plateaus.

    AbstractPosterCode

  2. Sci. Reports 2026

    Cross-Domain Generalization Failure in Lightweight Intrusion Detection Models for IIoT Networks

    MD Azizul Hakim, Md Shihab Uddin, Talha Ibne Anis

    Intrusion detectors trained on one industrial network fail on others without retraining. They lean on coarse port features, and the evaluation protocol alone can flip which network looks harder.

    arXiv

  3. Sci. Reports 2025

    Biologically inspired neural network layer with homeostatic regulation and adaptive repair mechanisms

    MD Azizul Hakim, Mohammad Ifazul Alam

    A neural layer that regulates its own activity and repairs itself after perturbation, evaluated on molecular and biomedical graph benchmarks.

    PaperCode

  4. IEEE QPAIN 2025

    A Novel Neuroplasticity-Based Approach for Enhanced Generalization in Computer Vision

    MD Azizul Hakim, Rashedul Arefin Ifty, Khaled Eabne Delowar, Samiul Azam Shuvo, Zarin Tasneem, Mohammad Ifazul Alam

    Layers with calcium-dependent regulation and homeostatic control reach 90.42% on CIFAR-10, 3.77 points above the baseline.

    Paper

  5. IEEE QPAIN 2025

    NutriGuard: LLM-Driven Nutritional Assessment for Chronic Disease Prevention

    MD Azizul Hakim, Rashedul Arefin Ifty, Khaled Eabne Delowar, Sazzad Hossen Chowdhury, Imdadur Rashid, Md Shakib

    OCR for English and Bengali food labels plus a fine-tuned Llama-3.2 that gives condition-aware dietary advice.

    Paper

  6. IEEE QPAIN 2025

    Neuroadaptive Intelligence: A Biologically-Inspired, Self-Repairing Neural Layer with Adaptive Learning for Molecular Graph Classification

    MD Azizul Hakim, Rashedul Arefin Ifty, Khaled Eabne Delowar, Samiul Azam Shuvo, Zarin Tasneem, Mohammad Ifazul Alam

    A self-repairing layer with adaptive learning, applied to molecular property prediction.

    Paper

  7. arXiv 2026preprint

    Brevity Constraints Reverse Performance Hierarchies in Language Models

    MD Azizul Hakim (sole author)

    Asking models to answer briefly can reverse which model families and sizes come out on top.

    arXiv

  8. arXiv 2026preprint

    Multi-Scale Temporal Homeostasis Enables Efficient and Robust Neural Networks

    MD Azizul Hakim (sole author)

    Coordinating homeostatic control across several timescales gives more stable training and better robustness.

    arXiv

Honors

Education