LAMAR: An Open Language-Aware Multilingual Alignment Reranker

Abstract

In multilingual retrieval a reranker will happily place a semantically equivalent document in the wrong language above one in the user's own — and the authors show existing multilingual rerankers do exactly that. Their open cross-encoder is trained to weigh language coherence alongside relevance: first distilling consistent relevance scores anchored in English, then applying preference alignment so same-language documents rank higher without losing semantic quality.

https://arxiv.org/abs/2607.22042