@inproceedings{wrobel-2025-unsupervised,
    title = "Unsupervised Detection of {LLM}-Generated {P}olish Text Using Perplexity Difference",
    author = "Wr{\'o}bel, Krzysztof",
    editor = "Kobyli{\'n}ski, {\L}ukasz  and
      Wr{\'o}blewska, Alina  and
      Ogrodniczuk, Maciej",
    booktitle = "Proceedings of the {P}ol{E}val 2025 Workshop",
    month = nov,
    year = "2025",
    address = "Warsaw",
    publisher = "Institute of Computer Science PAS and Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-paclic/2025.poleval-main.5/",
    pages = "26--38",
    abstract = "Inspired by zero-shot detection methods that compare perplexity across model pairs, we investigate whether computing perplexity differences on whole-text character-level perplexity can effectively detect LLM-generated Polish text. Unlike token-level ratio methods that require compatible tokenizers, our approach enables pairing any models regardless of tokenization. Through systematic evaluation of 91 model pairs on the PolEval 2025 {\'S}MIGIEL shared task, we identify Gemma-3-27B and PLLuM-12B as optimal, achieving 81.22{\%} accuracy on test data with unseen generators. Our difference-based approach outperforms token-level ratio methods (+5.5pp) and single-model baselines (+8.3pp) without using training labels, capturing asymmetric reactions where human text causes greater perplexity divergence than LLM text. We demonstrate that complementary model pairing (multilingual + monolingual) and architectural quality matter more than raw model size for this task."
}