Researchers have developed OmniAlign, a single model that can align words and sentences across languages. Unlike existing tools, which typically focus on one level of alignment (word or sentence), OmniAlign supports both with a lightweight design. The model uses an encoder-only backbone to induce word alignments from contextualized token similarity matrices and combines sentence embeddings with dynamic programming for document-level alignments. Experiments show that OmniAlign achieves competitive performance on both word- and sentence-alignment benchmarks, including unseen language pairs. Interestingly, fine-tuning the model on short texts can even improve alignment quality while retaining its ability to understand long contexts. --- Why it matters: OmniAlign matters because it provides a unified solution for cross-lingual sequence alignment, which is essential for building and exploiting parallel corpora in multilingual settings. This could simplify workflows for practitioners working with large datasets and multiple languages. Source: https://arxiv.org/abs/2608.18474