Mixed-language queries can improve multilingual retrieval performance, but the benefit depends on which language's documents you're searching; English dominance and linguistic similarity between languages predict when mixing helps.
This paper studies how dense retrieval systems handle mixed-language queries by blending embeddings from different language translations. Using controlled experiments on multilingual data, the researchers find that mixing languages often improves retrieval, but English dominates—mixing helps most when searching non-English documents, while pure English works best for English-heavy collections.