About
I'm Mohammad Zbeeb, a Computer Science and Engineering student at the American University of Beirut and a research assistant with Prof. Bernard Ghanem and Dr. Hasan Abed Al Kader Hammoud at KAUST's Image and Video Understanding Lab. My work centers on large language models, reasoning transfer, alignment, and safety for underrepresented languages.
You can reach me at mbz02@mail.aub.edu.
Current Research
- Reasoning Vectors (first author, ICLR 2026 submission) — extract a reasoning vector from RL-trained models and add it to other LLMs via task arithmetic to transfer chain-of-thought ability without retraining.
- Hala Technical Report (co-first, ACL 2026 submission) — Arabic-centric instruction and translation models built through translate-and-tune pipelines, SLERP merging, and bilingual supervision; reached #1 paper on Hugging Face.
- AraLingBench (co-first) — fully human-annotated benchmark evaluating the Arabic linguistic competence of 35 LLMs, revealing surface fluency vs. deeper grammatical gaps.
- Ongoing work on safety assets for low-resource languages and inference-time optimizations such as domain-specific draft models for speculative decoding.
Industry & Startups
- Edgebot: post-training team member applying literature-backed recipes to surpass internal LLM benchmarks for systematic trading.
- Chimera: led development of an AI-first cybersecurity engine that inspects network traffic with quantized, efficient multimodal models.
- Ubilite: explored differentiable optimization for digital circuit design.
Teaching & Community
- AUB teaching assistant for Intro to ML (3x) and AI in Medicine; delivered lectures, designed projects, and ran research paper sessions.
- Preprint: Optimizing Deep Neural Networks using Safety-Guided Self Compression — safety-driven quantization with preservation sets; selected top 10 in the IEEE Student Paper Contest.
- Co-founded AI Journey and helped launch the Blue Room, AUB's first AI lab with GPU access and remote compute for student research.
Interests
I'm currently thinking about on-policy reward generation from internal representations, activation-level methods for removing unsafe behavior, and data selection pipelines that prioritize medium-loss samples for stronger generalization.
Acknowledgments
My journey in research would not have been possible without the exceptional guidance and mentorship I've received.
Dr. Hasan Hammoud has been instrumental in shaping my understanding of machine learning research. His patient guidance through the intricacies of language models, his insights on rigorous experimentation, and his unwavering support have been transformative. From debugging late-night training runs to discussing the philosophical implications of every experiment, Hassan has taught me not just the technical craft of research, but the intellectual curiosity and rigor that defines great science.
I am deeply grateful for his mentorship, collaboration, and friendship throughout this journey.