I am a first-year PhD student at Saarland University and the German Research Center for AI (DFKI)One of the largest non-profit AI research institutions in the world, with its main site in Kaiserslautern and further sites across Germany, including Saarbrücken.. My research focuses on interpretability and efficiency in multilingual language models, supervised by Dr. Simon Ostermann and Prof. Dr. Josef van Genabith. For an overview of my work, see my Google Scholar profile or scroll further down this page. I am also contributing to the pre- and post-training of SOOFI, the first German-centric open-source foundation model, focusing on multilingual reasoning data generation.
Along the way, I have been fortunate to attend the ALPS NLP SchoolAdvanced Language Processing School, a week-long NLP school held in the French Alps, co-organized by Grenoble Alpes University and Naver Labs Europe. at Grenoble University and the MLSSMachine Learning Summer School, a long-running series of intensive graduate-level schools on statistical machine learning, running since 2002. summer school at Columbia University, as well as spend a month as a research visitor at KInITKempelen Institute of Intelligent Technologies, Slovakia's first independent private AI research institute, based in Bratislava. in Slovakia. I have also had the pleasure of co-organizing two events for the NLP community: TaCoSTagung der Computerlinguistik-Studierenden, Germany's annual student conference on computational linguistics., the 33rd Student Conference on Computational Linguistics, and the LT-Bridge Winter SchoolA DFKI-led project on multilingual and efficient NLP technologies, spanning chatbots, machine translation, and low-resource language processing. on parameter-efficient methods for NLP.
I completed my Master’s degree in Language Science and Technology at Saarland University, supported by the DAADDeutscher Akademischer Austauschdienst, Germany's national agency for funding international student and researcher exchange. and Erasmus+ scholarships. As part of my master's studies, I also spent a semester at the Center for Mind and Brain Sciences at the University of Trento (Italy). During this time, I worked as a Research Assistant in the Multilingual Language Technology Lab at DFKI, collaborating closely with my current supervisors.
Before moving to Germany, I earned my Bachelor's degree in Linguistics and Pedagogy at Shevchenko Transnistrian State University (Transnistria)A breakaway territory on the eastern bank of the Dniester, formally part of Moldova but functioning as a self-governing, internationally unrecognized state since 1990.. As part of my undergraduate studies, I completed an exchange semester at Murray State University (USA) through the fully funded Global UGRADGlobal Undergraduate Exchange Program, funded by the U.S. Department of State, bringing undergraduates from underrepresented countries to study for a semester in the US. program.
My early fascination with the intersection of mathematics, technology, and language naturally steered me toward Artificial Intelligence and later to Natural Language Processing. This journey has been both challenging and rewarding, and I am excited to continue exploring the inner workings of multilingual models and contributing to their development.
My research goal is to move from empirical steering methods toward a principled theory of controllable representations, using multilinguality as a lens for structure that stays invisible in English-centric analyses. I organize this around three pillars.
A rough story of what I've worked on so far during my PhD, in the order it happened.
It started with finding that some neurons inside multilingual models are language-specific: they can be located and manipulated, and they naturally cluster by language family, mostly in the later layers of the network.
read the paper →Neurons turned out to be a coarse tool. A cleaner picture emerged with vectors: language behaves like a linear direction in the representation space, and steering along it is a better way to control which language a model responds in, especially when applied in later layers.
read the paper →Digging into what "language" means at the representation level, script turned out to be its own linear axis that can be manipulated, gated by a small set of language-agnostic attention heads that causally control it, again concentrated in the later layers.
read the paper →Zooming out, these simple linear edits at the embedding level started to look like a real alternative to heavier adaptation methods (supervised fine-tuning, RL, or parameter-efficient tuning) for steering model behavior.
read the paper →The next question was whether these directions compose. It turns out that steering vectors can be combined linearly, as long as they are sufficiently orthogonal to each other.
paper coming soonPutting the pieces together, a working theory is forming: models first extract language-specific features and identify the input language, then reason in a shared cross-lingual space aligned with whichever language dominates, and finally differentiate the output back into the target language. This is where I want to take the next chapters of the PhD.
no dedicated paper yet