Querit-Reranker-4B is a multilingual cross-encoder reranker initialized from Qwen3-Embedding-4B and further trained with a data-centric, reranking-oriented pipeline. Rather than relying on backbone scale alone, the model first learns broad query-document relevance matching from large-scale ranking supervision and then adapts to target ranking distributions through synthetic-query mining with teacher scores as continuous soft labels.
Selected checkpoints from different data mixtures and training runs are further consolidated with spherical linear interpolation (SLERP), yielding a single deployable reranker without runtime ensembling overhead. By jointly encoding each query-document pair, Querit-Reranker-4B captures fine-grained relevance signals for second-stage ranking and achieves strong performance across multilingual and English retrieval benchmarks.
As of August 29, 2026, Querit-Reranker-4B achieves the best average score of 71.09 among publicly available models on the MTEB Multilingual v2 reranking tasks, averaged over six tasks.
If you find Querit-Reranker useful for your research or applications, please cite our paper:
bibtex
1@misc{zhong2026queritrerankertrainingcompactmultilingual,
2 title={Querit-Reranker: Training Compact Multilingual Rerankers via Efficient Label-Free Distribution Adaptation},
3 author={Yunfei Zhong and Jun Yang and Wei Huang and Yinqiong Cai and Haosheng Qian and Yixing Fan and Ruqing Zhang and Lixin Su and Daiting Shi and Jiafeng Guo},
4 year={2026},
5 eprint={2606.19037},
6 archivePrefix={arXiv},
7 primaryClass={cs.IR},
8 url={https://arxiv.org/abs/2606.19037},
9}
