@inproceedings{sahlen-etal-2026-miun,
    title = "{MIUN} {B}ias{P}atrol at {S}em{E}val-2026 Task 13: Why {TF}-{IDF} can Beat Transformers for {OOD} Code Detection",
    author = "Sahlen, Loviza  and
      Springfeldt, Thomas  and
      Fatima, Mehwish  and
      Shahzad, Raja Khurram",
    editor = "Kochmar, Ekaterina  and
      Ghosh, Debanjan  and
      North, Kai  and
      Komachi, Mamoru",
    booktitle = "Proceedings of the 20th {I}nternational {W}orkshop on {S}emantic {E}valuation (2026)",
    month = jul,
    year = "2026",
    address = "San Diego, California, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.semeval-1.312/",
    pages = "2469--2474",
    ISBN = "979-8-89176-414-9",
    abstract = "The increasing use of AI-generated code underscores the need for effective detection systems. However, their performance often deteriorates when faced with distribution shifts. This paper presents our system for SemEval-2026 Task 13: A, which focuses on binary classification of human-written versus machine-generated code across various programming languages and domains. We systematically compare traditional classifiers, such as Random Forest and XGBoost, which utilize statistical and TF-IDF features, against pipelines that leverage frozen embeddings from advanced code transformers like UniXcoder and GraphCodeBERT. Our results reveal a notable trade-off, i.e., while transformer-based pipelines excel in in-distribution validation (reaching up to 0.89 Macro F1), they experience severe performance drops up to 77{\%}; when applied to out-of-distribution languages and domains. In contrast, models employing TF-IDF with tree-based classifiers demonstrate significantly greater stability. We identify this fragility as a result of a bias toward superficial formatting, particularly whitespace, which is accentuated by transformers. By implementing simple space normalization, we enhance the out-of-distribution robustness of traditional models; however, this also highlights the ongoing dependence of embeddings on these non-semantic features. Our findings suggest that for creating generalizable code detection systems, straightforward, well-normalized lexical features may be more reliable than complex, unrefined embeddings."
}