Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

XProvence: Zero-Cost Multilingual Context Pruning for Retrieval-Augmented Generation

Youssef Mohamed, Mohamed Elhoseiny, Thibault Formal, Nadezhda Chirkova

TL;DR

RAG systems incur high cost from processing large retrieved contexts, especially in multilingual settings. XProvence extends the Provence framework to multilingual use by training a multilingual reranker to perform zero-cost context pruning, using cross-lingual transfer, data translation, and multilingual data annotation. Across four multilingual question-answering benchmarks, it achieves substantial context reduction with minimal to no degradation and outperforms strong baselines, while remaining resource-efficient. The work provides practical release of code and models for widespread deployment in multilingual RAG pipelines.

Abstract

This paper introduces XProvence, a multilingual zero-cost context pruning model for retrieval-augmented generation (RAG), trained on 16 languages and supporting 100+ languages through effective cross-lingual transfer. Motivated by the growing use of RAG systems across diverse languages, we explore several strategies to generalize the Provence framework-which first integrated efficient zero-cost context pruning directly into the re-ranking model-beyond English. Across four multilingual question answering benchmarks, we show how XProvence can prune RAG contexts with minimal-to-no performance degradation and outperforms strong baselines. Our model is available at https://huggingface.co/naver/xprovence-reranker-bgem3-v2.

XProvence: Zero-Cost Multilingual Context Pruning for Retrieval-Augmented Generation

TL;DR

RAG systems incur high cost from processing large retrieved contexts, especially in multilingual settings. XProvence extends the Provence framework to multilingual use by training a multilingual reranker to perform zero-cost context pruning, using cross-lingual transfer, data translation, and multilingual data annotation. Across four multilingual question-answering benchmarks, it achieves substantial context reduction with minimal to no degradation and outperforms strong baselines, while remaining resource-efficient. The work provides practical release of code and models for widespread deployment in multilingual RAG pipelines.

Abstract

This paper introduces XProvence, a multilingual zero-cost context pruning model for retrieval-augmented generation (RAG), trained on 16 languages and supporting 100+ languages through effective cross-lingual transfer. Motivated by the growing use of RAG systems across diverse languages, we explore several strategies to generalize the Provence framework-which first integrated efficient zero-cost context pruning directly into the re-ranking model-beyond English. Across four multilingual question answering benchmarks, we show how XProvence can prune RAG contexts with minimal-to-no performance degradation and outperforms strong baselines. Our model is available at https://huggingface.co/naver/xprovence-reranker-bgem3-v2.
Paper Structure (20 sections, 2 figures, 1 table)

This paper contains 20 sections, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Background
  3. Reranking.
  4. Provence.
  5. Training Methodology for XProvence
  6. Cross-Lingual Transfer.
  7. Data Translation.
  8. Multilingual Data Annotation.
  9. Experimental Setup
  10. Training details.
  11. Evaluation.
  12. Evaluation datasets.
  13. Baselines.
  14. Experiments
  15. RQ1: What is the best training recipe for XProvence?
  16. ...and 5 more sections

Figures (2)

  • Figure 1: XProvence speeds up generation in multilingual RAG pipelines through zero-cost context pruning, using an extra prediction head on a multilingual reranker to classify each sentence as relevant or irrelevant to the query.
  • Figure 2: Main results: for each dataset, we present a Pareto front obtained by varying the pruning threshold and averaged over languages. Lines located closer to the top right corner are best performing. Notation $L_{q}$/$L_{cntx}$ denotes the language of query/passage. For MKQA, we perform controlled experiments with various language settings, including settings $L_{cntx}=L_{q}$ and $L_{cntx}=En$. For the former setting, we present separately the results for languages seen and unseen in the training data. For the remaining datasets, we follow their original setting.