Querit-Reranker: Training Compact Multilingual Rerankers via Efficient Label-Free Distribution Adaptation

2026-06-17

AI Summary

This work addresses the challenge of cross-lingual and cross-domain generalization in multilingual reranking models, which typically rely on costly task-specific labeled data. The authors propose an efficient, label-free adaptation framework that first pretrains on large-scale ranking data, then leverages soft labels from a teacher model to synthesize queries for distributional adaptation. A single high-performance deployment model is obtained by fusing multi-task checkpoints via spherical linear interpolation. Combining a Mixture-of-Experts backbone with a cross-encoder architecture, the method achieves substantial gains in nDCG@10 on BEIR and MIRACL benchmarks and outperforms larger embedding-based baselines across multilingual reranking tasks in MTEB. Notably, Querit-Reranker-4B attains state-of-the-art performance among publicly available models.

Abstract

Deployable multilingual rerankers must generalize across languages, domains, and target ranking tasks while remaining efficient enough for second-stage reranking. However, adapting them to new target distributions typically requires extensive task-specific relevance annotations, which are costly to obtain. We present Querit-Reranker, a family of multilingual cross-encoder rerankers trained with a data-centric pipeline for label-efficient adaptation. We instantiate it as Querit-Reranker-A0.4B, initialized from an in-house MoE backbone with 0.4B activated parameters, and Querit-Reranker-4B, initialized from Qwen3-Embedding-4B. Our pipeline first learns general relevance modeling from large-scale ranking-oriented data, then adapts to target distributions through synthetic-query mining with teacher scores as continuous soft labels. To consolidate complementary task-adapted strengths, we further merge checkpoints via spherical linear interpolation, obtaining a single deployable model without runtime ensembling overhead. Using Qwen3-Embedding-0.6B as the shared first-stage retriever, Querit-Reranker-A0.4B improves average nDCG@10 from 54.11 to 59.28 on BEIR and from 59.87 to 67.70 on MIRACL. On MTEB Multilingual v2 Reranking, it also substantially outperforms larger embedding-based baselines, while Querit-Reranker-4B further achieves state-of-the-art performance among publicly available models. We release both models on Hugging Face.

Problem

• multilingual reranking
• label-efficient adaptation
• distribution adaptation
• cross-encoder
• second-stage reranking

Innovation

• label-free adaptation
• synthetic-query mining
• spherical linear interpolation
• multilingual reranking
• compact cross-encoder