Tag
The paper proposes query-aware routing for multilingual encoders, combining a query-only low-rank adapter (SampoTron) trained against frozen Nemotron-3-Embed-1B document embeddings with deterministic language routing. This yields a 20.9% relative nDCG@10 gain across six English, Finnish, and Swedish retrieval directions while preserving same-language performance and the existing document index.
This paper investigates the cause of cross-lingual retrieval asymmetry in multilingual embedding models. The authors propose and test the hub-mediation hypothesis, finding that hubness, not anisotropy, is the dominant cause, and recommend using CSLS instead of cosine similarity.