@inproceedings{xing-etal-2026-test,
    title = "Test-Time Adaptation of an Offline Multimodal Foundation Model for Simultaneous Speech Translation",
    author = "Xing, Yi  and
      Yu, Manli  and
      Liu, Pengfei  and
      Meng, Helen",
    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/bulk-corrections-2026-07-02/2026.iwslt-1.27/",
    doi = "10.18653/v1/2026.iwslt-1.27",
    pages = "238--246",
    ISBN = "979-8-89176-411-8",
    abstract = "End-to-end simultaneous speech-to-text translation (SimulST) systems typically rely on complex architectures and sophisticated training strategies. In contrast, we propose a simple approach that combines conventional pause-based segmentation for streaming audio input with a strong off-the-shelf multimodal foundation model adapted at test-time for translation. To achieve simultaneity, we adopt a variant of the classic wait-k read-write policy to control the interaction between audio input and translation output, and use a multi-turn conversation format with response prefilling and key-value caching for coherent translation and computational efficiency. Experiments on the official development sets of the IWSLT 2026 SimulST shared task show that our system achieves a better quality{--}latency trade-off than the cascaded baseline across all language directions and latency regimes, highlighting the effectiveness of this simple yet powerful approach."
}