PhD researcher · UCU / Research Engineer · Lapa LLM
I study what it means for large language models to forget, and how to build models that can.
I’m a PhD researcher at Ukrainian Catholic University and a Research Engineer at Lapa LLM. My work connects machine unlearning, mechanistic interpretability, information integrity, and responsible data systems.
Recent activity
I presented Between Suppression and Collapse: Evaluating Narrative Unlearning with LENS, co-authored with George Fletcher, at the 2nd Workshop on Machine Unlearning and Privacy Preservation in Naples. Presentation details
FairForget received a six-month allocation on JUPITER Booster at the Jülich Supercomputing Centre through the EuroHPC Development Access Call. I am the principal investigator for this allocation, which supports our research on fairness in machine unlearning. EuroHPC programme
With Tetiana Zakharchenko, I study how machine unlearning affects fairness and hidden biases. Programme announcement
I spoke at Align, Forget, Repeat in Lviv about whether LLMs can forget information while retaining their capabilities. Talk details
Memory, erasure, and responsibility in large language models.
My PhD research asks when an LLM can be said to know something, what changes when that knowledge is targeted for removal, and how forgetting can be verified rather than merely claimed. I study both the technical process of machine unlearning and the responsibilities that follow from deciding what a model should no longer retain. A central question is how mechanistic interpretability can inform unlearning.
Read the PhD proposalAcross my work
Related questions I pursue across research and teaching.
What models retain, how forgetting can be verified, and how mechanistic interpretability can inform unlearning.
Disinformation mitigation, narrative unlearning, and reducing bias through unlearning.
How people, institutions, and invisible entities enter data models.
2025–2026
2026
Open resources and collaborative technical projects.
I co-developed the Ukrainian translation and adaptation of StereoSet with Oleksandra Ostafiichuk and Tetiana Zakharchenko.
I participated with the DFS team in a challenge on recognizing Ukrainian handwritten documents. We placed fourth on the private leaderboard with a three-stage pipeline for layout detection, crop-specific transcription, and conservative OCR post-processing.
Notes
Working explanations of ideas behind the research.
Deleting a record from storage does not necessarily remove its effect from a trained model. It helps to separate four related ideas: erasure removes or disables access to an artifact; unlearning aims to remove or limit its influence on a model; exclusion concerns what is not collected, labeled, or included; and forgetting describes the wider technical and institutional process.
The practical problem is verification. Evaluation needs to establish how much targeted influence remains and what the tests can detect. It also needs a record of what changed, who made the changes, and how they affected the model’s remaining capabilities.
Related paperManipulation is often subtle at the level of one edit or message. A phrase can appear neutral in isolation while still participating in a repeated narrative across sources, channels, or time.
My work on Wikipedia edits and Telegram channels moves between these scales: detecting local textual signals, grouping related claims, and then examining the network through which a narrative is amplified.
Master's thesisA data model is not a neutral inventory. Decisions about entities, relationships, and levels of abstraction determine what a system can record and what remains invisible.
ONION makes these decisions visible through stakeholder participation. GARLIC adapts that approach for teaching, using role-play to help students examine whose needs a model represents and where it needs revision.
Related paperBackground
At Lapa LLM, I work on disinformation mitigation strategies, narrative unlearning, and bias mitigation through unlearning.
Lapa LLM project ↗I am a mentor in Responsible AI for Ukraine, a research programme connecting Ukrainian students with researchers through NYU and UCU.
Mentoring & supervision ↗I teach natural language processing at Ukrainian Catholic University, connecting applied tasks with research, ethics, and project-based assessment.
NLP course & materialsAs a visiting lecturer at American University Kyiv, I teach Web Enabled Decision Support Systems (SDT 520) in the Technology Leadership & AI master’s programme. I taught the course in spring 2025 and 2026.
AUK course detailsI completed my M.Sc. in Data Science with honors at UCU in June 2025 and began my PhD there in October 2025. I also hold a master’s degree in Radio Engineering from Kharkiv National University of Radio Electronics. Before beginning my PhD, I spent more than seven years in software engineering and technical leadership, including work at Artur’In.
Curriculum vitaeI’m especially open to research collaborations on LLM forgetting, mechanistic interpretability, and responsible AI. I also welcome speaking invitations, student projects, and other opportunities to work together.
Get in touch ↗George Fletcher, Julia Stoyanovich, and Tetiana Zakharchenko.
Research details