publications

2025

1. Understanding Verbatim Memorization in LLMs Through Circuit Discovery

   Ilya Lasy, Peter Knees, and Stefan Woltran

   In Proceedings of the First Workshop on Large Language Model Memorization (L2M2) , Aug 2025

   Link Abs Bib

   Underlying mechanisms of memorization in LLMs—the verbatim reproduction of training data—remain poorly understood. What exact part of the network decides to retrieve a token that we would consider as start of memorization sequence? How exactly is the models’ behaviour different when producing memorized sentence vs non-memorized? In this work we approach these questions from mechanistic interpretability standpoint by utilizing transformer circuits—the minimal computational subgraphs that perform specific functions within the model. Through carefully constructed contrastive datasets, we identify points where model generation diverges from memorized content and isolate the specific circuits responsible for two distinct aspects of memorization. We find that circuits that initiate memorization can also maintain it once started, while circuits that only maintain memorization cannot trigger its initiation. Intriguingly, memorization prevention mechanisms transfer robustly across different text domains, while memorization induction appears more context-dependent.

   @inproceedings{lasy-etal-2025-understanding,
     title = {Understanding Verbatim Memorization in {LLM}s Through Circuit Discovery},
     author = {Lasy, Ilya and Knees, Peter and Woltran, Stefan},
     editor = {Jia, Robin and Wallace, Eric and Huang, Yangsibo and Pimentel, Tiago and Maini, Pratyush and Dankers, Verna and Wei, Johnny and Lesci, Pietro},
     booktitle = {Proceedings of the First Workshop on Large Language Model Memorization (L2M2)},
     month = aug,
     year = {2025},
     address = {Vienna, Austria},
     publisher = {Association for Computational Linguistics},
     pages = {83--94},
   }

2. Guiding Generative Storytelling with Knowledge Graphs

   Zhijun Pan, Antonios Andronis, Eva Hayek, Oscar AP Wilkinson, Ilya Lasy, and 4 more authors

   May 2025

   arXiv Abs Bib

   Large Language Models (LLMs) have shown great potential in automated story generation, but challenges remain in maintaining long-form coherence and providing users with intuitive and effective control. Retrieval-Augmented Generation (RAG) has proven effective in reducing hallucinations in text generation; however, the use of structured data to support generative storytelling remains underexplored. This paper investigates how knowledge graphs (KGs) can enhance LLM-based storytelling by improving narrative quality and enabling user-driven modifications. We propose a KG-assisted storytelling pipeline and evaluate its effectiveness through a user study with 15 participants. Participants created their own story prompts, generated stories, and edited knowledge graphs to shape their narratives. Through quantitative and qualitative analysis, our findings demonstrate that knowledge graphs significantly enhance story quality in action-oriented and structured narratives within our system settings. Additionally, editing the knowledge graph increases users’ sense of control, making storytelling more engaging, interactive, and playful.

   @misc{pan2025guidinggenerativestorytellingknowledge,
     title = {Guiding Generative Storytelling with Knowledge Graphs},
     author = {Pan, Zhijun and Andronis, Antonios and Hayek, Eva and Wilkinson, Oscar AP and Lasy, Ilya and Parry, Annette and Gadney, Guy and Smith, Tim J. and Grierson, Mick},
     month = may,
     year = {2025},
     eprint = {2505.24803},
     archiveprefix = {arXiv},
     primaryclass = {cs.CL},
   }

2024

1. TU Wien at SemEval-2024 Task 6: Unifying Model-Agnostic and Model-Aware Techniques for Hallucination Detection

   Varvara Arzt, Mohammad Mahdi Azarbeik, Ilya Lasy, Tilman Kerl, and Gábor Recski

   In Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024) , Jun 2024

   Link Abs Bib

   This paper discusses challenges in Natural Language Generation (NLG), specifically addressing neural networks producing output that is fluent but incorrect, leading to “hallucinations”. The SHROOM shared task involves Large Language Models in various tasks, and our methodology employs both model-agnostic and model-aware approaches for hallucination detection. The limited availability of labeled training data is addressed through automatic label generation strategies. Model-agnostic methods include word alignment and fine-tuning a BERT-based pretrained model, while model-aware methods leverage separate classifiers trained on LLMs’ internal data (layer activations and attention values). Ensemble methods combine outputs through various techniques such as regression metamodels, voting, and probability fusion. Our best performing systems achieved an accuracy of 80.6% on the model-aware track and 81.7% on the model-agnostic track, ranking 3rd and 8th among all systems, respectively.

   @inproceedings{arzt-etal-2024-tu,
     title = {{TU} {W}ien at {S}em{E}val-2024 Task 6: Unifying Model-Agnostic and Model-Aware Techniques for Hallucination Detection},
     author = {Arzt, Varvara and Azarbeik, Mohammad Mahdi and Lasy, Ilya and Kerl, Tilman and Recski, G{\'a}bor},
     editor = {Ojha, Atul Kr. and Do{\u{g}}ru{\"o}z, A. Seza and Tayyar Madabushi, Harish and Da San Martino, Giovanni and Rosenthal, Sara and Ros{\'a}, Aiala},
     booktitle = {Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024)},
     month = jun,
     year = {2024},
     address = {Mexico City, Mexico},
     publisher = {Association for Computational Linguistics},
     doi = {10.18653/v1/2024.semeval-1.173},
     pages = {1183--1196},
   }