LAMAR Reranker Boosts Multilingual RAG by Prioritizing Query Language

In multilingual retrieval-augmented generation (RAG), retrievers often pull relevant documents in multiple languages. A reranker then orders these candidates before answer generation. However, researchers at Korea University found that existing multilingual rerankers do not consistently prioritize documents written in the same language as the query, even when semantically equivalent alternatives are available. This oversight can degrade answer quality, as the language of the source document affects how the generator processes the content. To solve this, the team introduces LAMAR (Language-Aware Multilingual Alignment Reranker), an open-source cross-encoder trained to account for both semantic relevance and language coherence.

LAMAR employs a two-stage training process. First, English-anchored relevance distillation ensures consistent relevance scoring across multilingual inputs by using English as a pivot language. Second, preference alignment for language coherence encourages documents in the same language as the query to receive higher rankings without sacrificing semantic relevance. In controlled experiments designed specifically to isolate language coherence, LAMAR achieved the best results overall and across all languages tested individually. It also remains competitive on established multilingual reranking benchmarks (e.g., Mr. TyDi). In practical first-stage retrieval settings, LAMAR outperformed baselines on all reported metrics. The model is available on arXiv and demonstrates that explicit language awareness can significantly improve multilingual RAG pipelines.