LAMAR reranks multilingual RAG results by relevance and language coherence
nlpai-lab · hf · 2026-07-27
- The paper studies multilingual RAG reranking and finds that existing rerankers do not reliably prefer documents in the same language as the query, even when semantically equivalent options exist.
- It releases LAMAR, a language-aware multilingual cross-encoder that balances semantic relevance with language coherence.
- Training combines English-anchored relevance distillation with preference alignment to encourage same-language documents without sacrificing relevance.
- In a controlled language-coherence experiment, LAMAR achieves the best overall performance across all languages tested.
- It also remains competitive on standard multilingual reranking benchmarks and achieves the best reported results in practical first-stage retrieval settings.
More from Research
- Paul David’s 106-page essay traces how open science emerged in the Scientific Revolution — michael_nielsen · 2026-07-27
- A note revisits why Newton’s second law is written as F = ma — michael_nielsen · 2026-07-27
- PRO-LONG scores 97.4% on ARC-AGI-3 with a log-file harness and 30-line prompt — srchvrs · 2026-07-27
- AI for biology splits into “AI scientists” and harder problems that need wet-lab validation — rishabh16_ · 2026-07-27
- SceneActBench tests whether VLM agents can act in full 3D scenes — Yifei Zhao · 2026-07-27
- Tencent Hunyuan maps scaling laws for native multimodal pretraining from scratch — Tencent-Hunyuan · 2026-07-27