Reproduce “LAMAR: An Open Language-Aware Multilingual Alignment Reranker” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Seongtae Hong, Youngjoon Jang, Jungseob Lee, Seungyoon Lee, Heuiseok Lim
Researchers at Korea University developed LAMAR, a 0.6B parameter open language-aware multilingual alignment reranker designed to jointly consider semantic relevance and language coherence for Retrieval-Augmented Generation (RAG) systems. The model achieved state-of-the-art performance in language-coherence evaluations (e.g., 0.9689 nDCG@1 on XQuAD) and competitive general multilingual reranking results (86.84 average nDCG@10 across five MTEB benchmarks).
Key insights
- Explicitly encouraging language coherence during reranking, in addition to semantic relevance, leads to a reranker that effectively prioritizes same-language documents among semantically equivalent candidates.
- A two-stage training approach, first distilling relevance and then aligning preferences, successfully balances general reranking effectiveness with specific language awareness.
- The language of the retrieved context directly impacts the quality of answers generated by LLMs, making language coherence a critical factor for multilingual RAG performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.