About
I am a PhD researcher at TU Wien, advised by Terra Blevins and Allan Hanbury. I am also currently a SPAR fellow, working on model diffing.
I work on how language models acquire natural language, how that knowledge is stored, and whether different types of knowledge (e.g. factual and linguistic) are stored in different ways, including in multilingual settings that go beyond high-resource languages.
Mechanistic interpretability · Model diffing · Multilingual NLP · Language modeling · AI safety
News
- Oct 2026 I presented our paper on word order preferences in language models as an oral at EMNLP 2026 in Budapest. Slides and paper are on the project page.
- Sep 2026 Global PIQA was accepted to NeurIPS 2026 as a poster.
- Sep 2026 I joined SPAR as a fellow, working on model diffing.
Education
- PhD in Computer Science, TU Wien, 2023 to expected autumn 2027. Thesis: “Factual and Linguistic Knowledge in Language Models”.
- MA in Byzantine and Modern Greek Studies, University of Vienna, 2022. Thesis.
- BA in Linguistics, Saint Petersburg State University, 2014.
Teaching
Since 2023 I have been teaching at TU Wien and the University of Klagenfurt, mostly courses on NLP, information retrieval and programming, with classes of up to 300 students. I also teach e-commerce and innovation, courses on the business side of digital technology.
Hobbies
I love learning languages: I speak Russian (native), German, English, Albanian and Modern Greek fluently, some French, and I am currently learning Indonesian. I also love almost any form of art, the world of perfume, spending time in nature, collecting mushrooms, and sailing. And running: I ran my first marathon in April 2026. :)
Contact
I'm always happy to collaborate on projects involving multilingual NLP and mechanistic interpretability, or just to chat about NLP and linguistics. Feel free to reach out.
varvara dot arzt at tuwien dot ac dot at