@inproceedings{gorantla-etal-2026-teamsls,
    title = "{T}eam{SLS} at {S}em{E}val-2026 Task 13: Detecting Machine-Generated Code with {C}ode{BERT} and Structural Features",
    author = "Gorantla, Sai Laasya  and
      Naveen, Shreemithra  and
      Bethard, Steven",
    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-1.318/",
    pages = "2520--2526",
    ISBN = "979-8-89176-414-9",
    abstract = "We describe our system for SemEval-2026 Task 13 Subtask A, which focuses on detecting whether source code is written by a human or generated by an AI system. We propose a hybrid approach that combines semantic embeddings from CodeBERT with lightweight, language-agnostic structural features extracted using Tree-sitter. We compute normalized structural ratios such as nesting depth, logic density, complexity per line, average line length, and punctuation frequency. These structural signals are concatenated with CodeBERT embeddings and passed to a linear classifier for binary prediction. Experimental results on the official validation split show that combining semantic and normalized structural representations substantially improves the model{'}s detection performance on seen-language distributions. However, results on unseen test data reveal significant performance degradation under cross-language distribution shifts. On the official leaderboard, our system ranked 47th out of 81 participating teams."
}