creature 1

Ali Holmov

creature 1
creature 2 About
Ali

Hey, I'm Ali, a CS master's student at TUM. Currently, I'm interested in mechanistic interpretability, and the question that excites me most is how deep learning models represent knowledge. I also like building things. Right now I'm working part-time at Helmit on ML models for online child safety.

Before that, I did research projects at EML Munich (user models, belief extraction), hessian.AI (knowledge editing) and King's College London (AI safety), and studied CS at Marburg & philosophy at Saint Petersburg State University. When I'm not working, I like to swim and fence, read (mostly fiction & poetry), and do jigsaw puzzles or sudoku. I also have a thing for medieval bestiary art, which explains the creatures on this site.

P.S. I dropped some letters from my surname (Kholmovaia) in 2026 and now use the simplified version.

creature 3 Publications
User Model Extraction via Belief Self-Distillation
A. Holmov, Y. Huang, K. Bykov, Z. Akata. Preprint, 2026. Code
One Mask to Rule Them All: On Hidden Facts after Editing and How to Find Them
A. Holmov, P. Youssef, N. Schoots, C. Seifert. Findings of ACL 2026. Code
The Propensity for Density in Feed-forward Models
N. Schoots, A. Jackson, A. Kholmovaia, P. McBurney, M. Shanahan. ECAI 2024.
Physicians Trust AI Less in Complex Cases, but Rely on It Just as Much
N. Schlicker, N. Steffens, A. Kholmovaia, O. Loukas, L. E. Melms, M. C. Hirsch, M. Langer. Technology, Mind, and Behavior. 2026.
creature 4 Thoughts
Who do you think I am?
A visual essay on the beliefs language models form about their users. 2026.