@inproceedings{c-etal-2026-spycomet,
    title = "{S}py{C}omet at {S}em{E}val-2026 Task 11: When Adversarial Debiasing Backfires - A Comparison of Data-Level and Model-Level Debiasing",
    author = "C, Sai Aravind  and
      Saumya, Sunil  and
      Reddy, C Pothan",
    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-1.317/",
    pages = "2512--2519",
    ISBN = "979-8-89176-414-9",
    abstract = "We describe MLA-CI (Multi-Layer Adversarial for Content Invariance), a DeBERTa-v3-base system for SemEval-2026 Task 11 Subtask 1 on content-invariant syllogistic reasoning. MLA-CI combines multi-layer feature extraction, gradient-reversal adversarial training, structure-preserving template augmentation, implausible-class oversampling, and focal loss. Our principal contribution is a systematic ablation study, confirmed across three random seeds, showing that adversarial training at standard strength is counterproductive: removing gradient reversal improves the mean validation score from 26.41 {\ensuremath{\pm}} 0.99 to 38.15 {\ensuremath{\pm}} 5.32. Per-condition analysis reveals that gradient reversal over-suppresses plausibility-correlated features, creating an inverted content effect that disproportionately harms plausible-valid accuracy. A sweep over seven adversarial pressure values reveal that only very light adversarial pressure value ({\ensuremath{\leq}} 0.1) preserves accuracy, while the submitted adversarial pressure value (1.0 or above) cause severe degradation. We conclude that data-level debiasing through structure-preserving augmentation is more effective and robust than model-level adversarial debiasing for this task."
}