@inproceedings{chen-etal-2026-polymorphic,
    title = "Polymorphic Universal Transformer",
    author = "Chen, Yilong  and
      Gao, Zitian  and
      Xiao, Yihao  and
      Liu, Jason Klein  and
      Yang, Xinyu  and
      Luo, Yifan  and
      Luo, Haoming  and
      Ye, Zhengmao  and
      Liu, Tingwen  and
      Tao, Ran  and
      Dai, Bryan",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl/2026.acl-long.1809/",
    pages = "39001--39013",
    ISBN = "979-8-89176-390-6",
    abstract = "Although the Universal Transformer (UT) mitigates the diminishing returns of standard LLM scaling by decoupling parameter count from depth, it remains constrained by linear computational costs and rigid weight-sharing mechanisms. These limitations lead to severe functional homogeneity, which subsequently induces over-smoothing, representation rank collapse, and degraded reasoning performance. In this work, we present the first systematic study of Compute Distribution Skew, identifying it as the primary driver of extrapolation failure. This is a pathological phenomenon in ultra-deep recurrent Transformers characterized by a disproportionate distribution of contributions across recurrent steps, resulting in distinct functional states during prefix and suffix processing phases. To address this challenge, we propose the Polymorphic Transformer, which aims to achieve functional polymorphism and depth sparsity within a shared-parameter framework. By integrating conditional sparse subspaces, SiLU Attention, and an uncertainty-aware depth scheduler, our architecture mitigates power-method collapse and effectively decouples logical depth from computational cost. Experiments demonstrate that our model significantly enhances representation rank and robustness, achieving complex reasoning performance comparable to baseline while reducing computation by 64.7{\%}."
}