arXiv 2601.21853

LEMUR: Learned Multi-Vector Retrieval

By Elias Jääsaari, Ville Hyvönen, et al.

Published 2026-01-29

Citation lineage

Review the prior work and downstream research connected to this paper.

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…

View the original paper on arXiv