@inproceedings{sun-etal-2026-cuhksz,
    title = "The {CUHKSZ} System for the {IWSLT} 2026 Low-Resource Speech-to-Text Task",
    author = "SUN, ruiyan  and
      Li, Qingming  and
      Nakamura, Satoshi",
    editor = "Salesky, Elizabeth  and
      Anastasopoulos, Antonios  and
      Negri, Matteo  and
      Federico, Marcello",
    booktitle = "Proceedings of the 23rd International Conference on Spoken Language Translation ({IWSLT} 2026)",
    month = jul,
    year = "2026",
    address = "San Diego, USA (in-person and online)",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/corrections-2026-06/2026.iwslt-1.33/",
    doi = "10.18653/v1/2026.iwslt-1.33",
    pages = "296--304",
    ISBN = "979-8-89176-411-8",
    abstract = "This paper describes the CUHKSZ system for the IWSLT 2026 Low-Resource Speech-to-Text task. We propose Gradient-Driven Parameter Sharing (GDPS), a framework that analyzes inter-language gradient behaviors to automatically determine optimal language groupings and shared-private parameter ratios. Built upon SeamlessM4T-Medium, GDPS reduces negative transfer by specializing Layer 11 FFN2 while maintaining shared encoder representations across languages. Additionally, we incorporate curriculum distillation with progressive pseudo-label mixing and test-time reranking combining prior-BLEU weighting and self-consistency scoring. Evaluation on eight low-resource languages (bem, ckb, gle, hau, ibo, yor, aeb, est) demonstrates strongest gains on bem (+2.07 BLEU), hau (+1.50), and ibo (+0.38) compared to unified fine-tuning, while ckb and yor benefit more from prior-based reranking at inference."
}