Hugging Face has introduced LAMAR (Language-Aware Multilingual Alignment Reranker), a novel cross-encoder designed to improve multilingual retrieval-augmented generation (RAG). The model addresses a key limitation in existing multilingual rerankers: they often fail to prioritize documents written in the same language as the query, even when semantically equivalent documents exist across languages.
LAMAR employs a two-stage training process. First, it uses English-anchored relevance distillation to ensure consistent relevance scoring across multilingual inputs. Second, it applies preference alignment for language coherence, encouraging documents in the query language to rank higher without sacrificing semantic relevance.
In controlled experiments measuring language coherence, LAMAR achieved the best overall performance and across all individual languages tested. It also remained competitive on established multilingual reranking benchmarks. In practical first-stage retrieval settings, LAMAR outperformed other models across all reported metrics.
The release of LAMAR as an open model aims to advance research in multilingual information retrieval, particularly for RAG systems where document language can significantly impact answer generation quality.