Mechanistic Interpretability Researcher
I am a postdoctoral researcher in the Explainable Machine Learning group at TUM and Helmholtz Munich, led by Prof. Zeynep Akata. I work on mechanistic interpretability of vision and language models, with a focus on AI safety and alignment. I completed my Ph.D. at TU Berlin (summa cum laude), following an M.Sc. at TU Berlin and TU Eindhoven and a B.Sc. at Saint Petersburg State University.
I study how mechanistic understanding can support scalable AI control. My current work targets safety-relevant LLM states, with the longer-term goal of turning mechanistic signals into runtime monitoring and intervention tools for autonomous agents and interacting agent populations.
arXiv preprint arXiv:2609.31603
LLMs form a picture of the users they interact with. We propose Belief Self-Distillation, an approach for extracting and causally editing these internal user models. We show that changing inferred user intent alters refusal, even when the request stays fixed.
arXiv preprint arXiv:2605.09159
We trace LLM reasoning through persona directions, using these interpretable signals to predict answer correctness and explore interventions during generation.
Advances in Neural Information Processing Systems 38, 82938-82974
PRISM captures the multiple concepts encoded by LLM features rather than compressing them into a single explanation. We evaluate these multi-concept descriptions to make explanation quality and polysemanticity measurable.
Advances in Neural Information Processing Systems 38, 166757--166794
Feature visualizations can be manipulated without substantially changing model performance. We demonstrate this vulnerability with Gradient Slingshots and evaluate a defense for more reliable model auditing.
Machine Learning: Earth 1 (1), 015002
Climate teleconnections provide useful structure for subseasonal European weather forecasts. We incorporate these remote atmospheric signals into deep learning models to improve predictions and examine the physical drivers of weather regimes.
Advances in Neural Information Processing Systems 37
CoSy tests whether a neuron’s textual explanation matches its behavior. We generate examples of the described concepts and use the neuron’s activation responses to evaluate explanation quality.
Communications in Computer and Information Science 1947, 426--434
ImageNet models can rely on watermarks as shortcuts across multiple classes. We locate the affected representations and show how targeted changes can reduce reliance on these misleading cues.
Advances in Neural Information Processing Systems 36, 24804-24828
INVERT connects internal representations with interpretable concepts without requiring segmentation masks. The method makes concept discovery scalable while measuring conceptual alignment and statistical significance.
Transactions on Machine Learning Research
Unusual representations can point to hidden artifacts in a model. With DORA, we find these outliers without a reference dataset and show how they expose spurious features in vision models.
Transactions on Machine Learning Research
Bayesian neural networks learn different representations across posterior samples. We visualize and measure this diversity to study its relationship with predictive uncertainty.
World Conference on Explainable Artificial Intelligence, 549-572
Comparing functional and semantic relationships can uncover hidden shortcuts. Our Function-Semantic Contrast Analysis uses this contrast to support scalable bias discovery with less manual inspection.
Proceedings of the AAAI Conference on Artificial Intelligence 36 (6), 6132-6140
Adding noise to model weights can improve the explanations we extract. NoiseGrad uses this idea to enhance both local attribution maps and global feature visualizations with a simple, method-agnostic procedure.
Transactions on Machine Learning Research
A distribution of explanations reveals more than a single attribution map. We extend attribution methods to Bayesian neural networks to study uncertainty and variation in the evidence behind a prediction.
Technical University of Munich · Munich, Germany
Research in mechanistic interpretability for modern deep networks, with a focus on representations, feature reliability, and explanation methods for vision and language models.
Technische Universität Berlin · Berlin, Germany
Advisors: Prof. Marina Höhne and Prof. Klaus-Robert Müller. Thesis: “Explaining Representations in Deep Neural Networks.” Summa cum laude, 1.0.
TU Berlin and TU Eindhoven · Berlin, Germany / Eindhoven, the Netherlands
Double-degree program. EIT Digital Excellence Scholarship. Cum laude, 1.3.
Saint Petersburg State University · Saint Petersburg, Russia
Faculty of Mathematics and Mechanics, Department of Statistical Modelling. Grade: 4.47 / 5.
Lyceum 239 · Saint Petersburg, Russia