LAMAR: An Open Language-Aware Multilingual Alignment Reranker

LAMAR targets a gap in multilingual retrieval-augmented generation: existing rerankers may rank semantically equivalent documents without consistently favoring the query language, which can affect answer generation. It is an open language-aware multilingual cross-encoder. The method first uses English-anchored relevance distillation to establish cross-lingual relevance scores, then applies preference alignment to promote language coherence while preserving semantic relevance. According to the paper, LAMAR performs best in a controlled language-coherence experiment across all examined languages, remains competitive on standard multilingual reranking benchmarks, and leads on all reported metrics in practical first-stage retrieval settings.