Dumitru Verșebeniuc

M.Sc. candidate in Artificial Intelligence

My research sits at the intersection of mathematical reasoning, RL post-training, and multi-agent systems — with a core question: when do language models genuinely reason, and when do they simply retrieve memorized patterns?

Maastricht, the Netherlandsd.versebeniuc@dikaver.com

GitHub LinkedIn Google Scholar ORCID

Publications

Apr · 2026MDPI, Electronics

On Memorization and Generalization in Compact Transformers

byA. Härmä, A. Al-Saeedi, A. Changalidis, D. Verșebeniuc, M. Pietrasik, A. Wilbik

Paper

Feb · 2025Preprint

From Pattern Recognition to Reasoning: A Survey for Transformer Length Generalization in Algorithmic Tasks

byD. Verșebeniuc

supervised byA. Härmä

Paper

Nov · 2024BNAIC/BeNeLearn 2024 · Post-proceedings in Springer CCIS · Oral Session

Generative AI-Based Virtual Assistant Using Retrieval-Augmented Generation: An Evaluation Study for Bachelor Projects

byD. Verșebeniuc, M. Elands, S. Falahatkar, C. Magrone, M. Falah

supervised byM. Boussé, A. Härmä

Paper Code Demo Presentation

Jun · 2023B.Sc. Thesis, Maastricht University

Model-Based Clustering Multivariate EMA Time-Series Data

byD. Verșebeniuc

supervised byJ. Spanakis, M. Ntekouli

Paper Code

Education

M.Sc. Artificial Intelligence

Maastricht University

Sep 2023 – Present

B.Sc. Data Science & Artificial Intelligence

Maastricht University

Sep 2020 – Jun 2023