LAMAR: An Open Language-Aware Multilingual Alignment Reranker

Researchers introduced LAMAR, a language-aware multilingual cross-encoder reranker for search and RAG.

It combines semantic relevance with language coherence, promoting documents in the query language when meanings are similar.

Training uses English-anchored relevance distillation and preference alignment to balance matching and language choice.

LAMAR stays competitive on standard benchmarks while improving same-language ranking, which can help multilingual retrieval quality.