# Dumitru Verșebeniuc

> Dumitru Verșebeniuc is an AI security researcher and engineer building evaluations and tools to detect reward hacking and make AI systems trustworthy by design.

Personal research portfolio: https://dikaver.com
Also written as Dumitru Versebeniuc. Based in Chișinău, Moldova.

I build evaluations and research tools to understand how AI systems behave. My goal is to make them trustworthy and secure by design. My background spans research, full-stack engineering, and entrepreneurship.

My current focus is detecting reward hacking: when an agent earns reward without doing the intended task.

I’m open to research collaborations and PhD opportunities in AI security.

## Research interests

- Reward hacking (Current focus): Detecting agents that exploit an evaluation or reward signal instead of completing the intended task.
- AI control (Exploring): Repeatedly testing whether safety protocols hold up when an agent deliberately tries to subvert them.
- Scientist AI (Exploring): Exploring non-agentic systems that explain evidence, express uncertainty, and assess risk as guardrails.

## Publications

- [Optimization of Inverter Placement and Cable Routing for Distributed Inverter Topologies in Solar Plants](https://dikaver.com/publications/optimization-inverter-placement-cable-routing): M.Sc. Thesis, Maastricht University, 2026. Master's thesis on joint inverter placement and cable routing for solar plants, using mixed-integer programming and a four-stage decomposition.
- [On Memorization and Generalization in Compact Transformers](https://dikaver.com/publications/memorization-generalization-compact-transformers): MDPI, Electronics, 2026. Experiments on memory capacity, domain knowledge, and in-context abstraction in compact transformers, with implications for on-edge AI architectures.
- [From Pattern Recognition to Reasoning: A Survey for Transformer Length Generalization in Algorithmic Tasks](https://dikaver.com/publications/from-pattern-recognition-to-reasoning): Preprint, 2025. A survey of transformer length generalization in algorithmic tasks, comparing positional encodings, data formats, and compositional reasoning.
- [Generative AI-Based Virtual Assistant Using Retrieval-Augmented Generation: An Evaluation Study for Bachelor Projects](https://dikaver.com/publications/generative-ai-virtual-assistant-rag): BNAIC/BeNeLearn 2024 · Post-proceedings in Springer CCIS · Oral Session, 2024. An evaluation of a retrieval-augmented virtual assistant that helps Maastricht University students navigate bachelor project regulations.
- [Model-Based Clustering Multivariate EMA Time-Series Data](https://dikaver.com/publications/model-based-clustering-ema-time-series): B.Sc. Thesis, Maastricht University, 2023. Bachelor's thesis comparing model-based clustering methods for multivariate ecological momentary assessment data, including Lasso and VAR models.

## Education

- M.Sc. Artificial Intelligence, Maastricht University (Sep 2023 – Aug 2026).
- B.Sc. Data Science & Artificial Intelligence, Maastricht University (Sep 2020 – Jun 2023).

## Profiles and contact

- [ORCID](https://orcid.org/0009-0004-4660-9636)
- [Google Scholar](https://scholar.google.com/citations?user=G3XnPDYAAAAJ)
- [GitHub](https://github.com/DikaVer)
- [LinkedIn](https://www.linkedin.com/in/dima-vers/)
- [Email](mailto:d.versebeniuc@dikaver.com)