arXiv 2601.21853
LEMUR: Learned Multi-Vector Retrieval
By Elias Jääsaari, Ville Hyvönen, et al.
Published 2026-01-29
Mindmap
Browse the paper's core ideas, clusters, and relationships in a structured outline.
Multi-vector representations generated by late interaction models, such as ColBERT, enable superior retrieval quality compared to single-vector representations in information retrieval applications. In multi-vector retrieval systems, both queries and documents are encoded using one embedding per token, and similarity between queries and documents is measured by the MaxSim similarity measure. However, the improved qu…