Evgenii Kortukov, Piotr Komorowski, Florian Klein, Paula Engl, Gabriele Sarti, Seong Joon Oh, Sebastian Lapuschkin, Wojciech Samek
ICML Workshop Mechanistic Interpretability, 2026Reasoning models represent distributions over future behaviors; probing these lets us steer the outcome by selecting reasoning sentences.
Raghu Arghal, Phoebe Chen, Niall Dalton, Evgenii Kortukov, Calum McNamara, Angelos Nalmpantis, Moksh Nirvaan, Gabriele Sarti, Mario Giulianelli
ICML, 2026We propose a framework for evaluating goal-directedness in LLM agents, integrating behavioural evaluation with interpretability analyses of internal representations.
Alexander Panfilov, Evgenii Kortukov, Kristina Nikolić, Matthias Bethge, Sebastian Lapuschkin, Wojciech Samek, Ameya Prabhu, Maksym Andriushchenko, Jonas Geiping
ICLR, 2026LLMs dishonesty breaks jailbreak evaluations but activation probes can catch it.
Egor Zverev, Evgenii Kortukov, Alexander Panfilov, Soroush Tabesh, Sebastian Lapuschkin, Wojciech Samek, Christoph H. Lampert
ICLR, 2026We introduce an architectural change to LLMs that separates instructions from data, to improve their security.
Elisa Nguyen, Johannes Bertram, Evgenii Kortukov, Jean Y. Song, Seong Joon Oh
ArXiv, 2024Making XAI more useful by first asking the users about their needs, with a focus on Training Data Attribution.
Evgenii Kortukov, Alexander Rubinstein, Elisa Nguyen, Seong Joon Oh
COLM, 2024Studying context-memory knowledge conflicts as they appear in practice: how does factual knowledge influence LLM reading behaviors?
Elisa Nguyen, Evgenii Kortukov, Jean Song, Seong Joon Oh
NeurIPS Workshop XAIA, 2023Interviewing ML practitioners to explore the human factor of training data attribution explanations.
Saeed Ghoorchian, Evgenii Kortukov, Setareh Maghsudi
IEEE Open Journal of Signal Processing, 2023Utilizing random projections to apply an MAB-algorithm to a recommender system scenario, where the data is high-dimensional and user preferences change over time.
Saeed Ghoorchian, Evgenii Kortukov, Setareh Maghsudi
IEEE Open Journal of Signal Processing, 2023We study contextual multi-armed bandit problem where the features are costly and the agent has to simultaneously learn the reward distributions and the feature importances. The environment undergoes distribution shifts making the problem more challenging.