AIAny
Icon for item

LAMAR: An Open Language-Aware Multilingual Alignment Reranker

Reranks multilingual retrieval candidates to favour documents that are both semantically relevant and written in the same language as the query, using English-anchored relevance distillation and preference alignment; excels in language-coherence tests while remaining competitive on standard multilingual reranking benchmarks.

Introduction

Why this matters Most multilingual retrieval pipelines treat semantic relevance and document language as independent signals, which can produce answers in a different language than the user query—problematic for downstream generation in RAG systems. LAMAR focuses on the often-overlooked objective of language coherence: given semantically equivalent documents across languages, a reranker should prefer the one in the query language without sacrificing relevance.

Key Findings
  • English-anchored relevance distillation: uses English-grounded scores to make relevance assessments consistent across multilingual inputs, reducing cross-language scoring drift.
  • Preference alignment for language coherence: learns to tilt rankings toward documents that match the query language while retaining semantic relevance, producing more language-consistent top results.
  • Empirical performance: in a controlled language-coherence experiment LAMAR achieves the best overall and per-language results; on established multilingual reranking benchmarks it remains competitive, and in practical first-stage retrieval settings it improves end-to-end retrieval metrics when used to rerank candidates.
Who it's for and trade-offs

Great fit if you build multilingual RAG or search systems where answer language matters (user-facing QA, multilingual assistants, cross-lingual search) and you need a reranker that enforces language coherence without major drops in relevance. Look elsewhere if your pipeline requires purely language-agnostic aggregation (e.g., cross-lingual pooling where language diversity is preferred) or if latency/compute constraints cannot accommodate a cross-encoder reranker.

Method sketch

The model is a multilingual cross-encoder trained to combine semantic relevance and language-preference signals. Training uses an English-anchored teacher to harmonize relevance scores across languages, plus an alignment objective that biases rankings toward query-language documents. The paper evaluates both controlled language-coherence settings and standard multilingual reranking benchmarks to show the trade-offs between coherence and raw relevance.

Information

  • Websitearxiv.org
  • AuthorsSeongtae Hong, Youngjoon Jang, Jungseob Lee, Seungyoon Lee, Heuiseok Lim
  • Published date2026/07/24

More Items

Empirically studies how transformer-based native multimodal pre-training scales under fixed compute, deriving compute- and data-allocation power laws and an efficiency frontier for model size, token count, and data mixture; evaluates cross-modal transfer and multimodal in-context learning.

Studies train-time knowledge injection via hypernetworks that generate fixed LoRA adapters from large fact corpora, empirically characterizing power-law scaling across hypernetwork depth, width, and target model size and reporting improved OOD generalization.

Models long-horizon interactive literary simulation where characters and world co-evolve; introduces an open‑schema framework with a Character Agent and an LLM-based World Model, plus seven trainable tasks and a dataset from 57 books for benchmarking persistent narrative state.