@inproceedings{jumakhan-etal-2026-ajman,
    title = "Ajman University at {S}em{E}val-2026 Task 2: Overcoming Scale Collapse in Temporal Emotion Modeling via Residual Learning",
    author = "Jumakhan, Haseebullah  and
      Assad, Soud  and
      Abdullah, Seyed  and
      Al-Ayyoub, Mahmoud",
    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-1.66/",
    pages = "457--462",
    ISBN = "979-8-89176-414-9",
    abstract = "Ajman University Team develops a set of specialized architectures for longitudinal affective forecasting for SemEval-2026 Task 2. We establish a baseline for our performance with a standard transformer model that sets our performance floor in Subtask 1 (ranked 18). In Subtask 2A (ranked 7) and Subtask 2B (ranked 8), our main contribution is to address the problem of scale collapse. To address the scale collapse, we use a novel ``bifurcated leviathan'' architecture to combine residual learning with target scaling. Our additional contribution is that we counteract the effects of regression to the mean by using optimized covariance via specialized objective functions (CCC and Huber). We use these objective functions while enforcing strict user level data splits. Finally, we show empirically that standard gradient stabilization methods decrease zero shot cross subject generalization, even when they optimize intra subject memorization."
}