TU Wien CAIML

Overview News Researchers & Projects Team Partners

Ilya Lasy

ilya-lasy.webp

On This Page

Short Bio

Ilya Lasy is a PhD student and university assistant at the Databases and Artificial Intelligence Unit at TU Wien.

He holds an MS degree in Computer Science (2022) from Vilnius University, and a BS degree in Software Engineering (2020) from the Belarusian State University of Informatics and Radioelectronics. His Master’s thesis focused on neuro-symbolic approaches to dialogue systems. During his Master studies he was selected to attend the Eastern European Machine Learning Summer School, where he presented a poster based on his Master’s thesis. Besides academia, Ilya works part-time as a Machine Learning Engineer at Charisma.ai, building Large Language Models for the creative entertainment industry. His main research interests are interpretability of LLMs, disentanglement of knowledge representations and neuro-symbolic AI.

PhD Project - Towards Explainable and Knowledge-Driven Large Language Models

Supervised by Peter Knees and Stefan Woltran

Main focus of the project is to investigate the inner workings of Large Language Models to understand their learned representations and inference processes. By applying mechanistic interpretability methods—such as circuit discovery and sparse autoencoders — Ilya aims to “reverse engineer” current architectures to reveal how they process and generate information. His current work explores hallucination (generation of false information) and memorization (verbatim repetition), seeking to disentangle these processes from genuine generalization. This understanding could lead to mitigating these issues in future architectures, ultimately building more trustworthy AI systems. Additionally, Ilya investigates whether aforementioned knowledge-related processes could be substituted with symbolic AI components (like databases), creating hybrid systems that are more interpretable, efficient, and safe while maintaining high performance.

Publications and Conferences

  • Arzt, V., Azarbeik, M. M., Lasy, I., Kerl, T., & Recski, G. (2024). TU Wien at SemEval-2024 Task 6: Unifying Model-Agnostic and Model-Aware Techniques for Hallucination Detection. Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024) (pp. 1183–1196). Association for Computational Linguistics. doi:10.18653/v1/2024.semeval-1.173
  • Lasy, I. and Marcinkevičius, V. (2022) “Dialogue System Augmented with Commonsense Knowledge”, Vilnius University Open Series, pp. 68–76. doi:10.15388/LMITT.2022.7.