Chunyi Peng1,3, Zhipeng Xu1, Zhenghao Liu1, Yishan Li3, Yukun Yan2, Yu Gu1 Minghe Yu1 Ge Yu1 Maosong Sun2
1Northeastern University, 2Tsinghua University, 3ModleBest Inc.
News
26.04.28 Our Work is now accepted by SIGIR 2026🎉!
25.08.22 We upload MoRE-3B.
Environment
For training, answer generation, and evaluation processes:
conda create -n router python=3.11
conda activate router
pip install requirements_router.txt
For retriever and corpus construction processes:
conda create -n retriever python=3.11
conda activate retriever
pip install requirements_retriever.txt
Corpora Construction
For the text corpus, you can download enwiki-20241020 from Huggingface. Then preprocess, and index it with the following commands:
7z x enwiki-20241020-pages-articles-multistream.xml.zip.001
conda activate retriever
wikiextractor enwiki-20241020-pages-articles-multistream.xml.bz2 -o wiki_extracted
python wiki_preprocess.py
For the image corpus, you can directly download M-BEIR. To embed and index it, you can follow the repository
For the table corpus, you can download, embed and index Open-WikiTable following the repository, or you can download directly the one we have already preprocessed from here.
Retrievers Preparation
For the Text-Image Retriever, you can directly download UniIR
For the Table Retriever, you can train it with the help of repository, or you can download it directly from here.
Datasets
We have prepared all the text datasets in ./datasets, for images you need to download them from:
InfoSeek:InfoSeek images can be downloaded from OVENDyn-VQA:Dynamic VQA images can be downloaded from DynVQA_en.202412WebQA:WebQA images can be downloaded from Google Drive
Training
If you do not want to train the model, you can download R1-Router and skip this section to Evaluation
Data Synthesis
If you want to use the ready-to-use synthetic data directly, you can skip this section to Step-GRPO Training
First, we need to synthesis the data step by step:
bash src/data_synthesis/data_synthesis.sh
Step-GRPO Training
Our training framework is based on EasyR1, only you need to do is to download it and replace some files with the files in ./Easy-R1.
Then start training with the command:
conda activate router
bash examples/run_qwen2_5_vl_7b_stepgrpo.sh
Evaluation
We provide the evaluation pipeline for the R1-Router:
bash evaluation.sh
or, you can just evaluate the results we have provided by:
conda activate router
cd src
python evaluate.py --dataset_name all --method "r1-router3"
Acknowledgement
Our work is built on the following codebases, and we are deeply grateful for their contributions.
Citation
We appreciate your citations if you find our paper relevant and useful to your research!
@article{peng2025r1,
title={Learning to Route Queries across Knowledge Bases for Step-wise Retrieval-Augmented Reasoning},
author={Peng, Chunyi and Xu, Zhipeng and Liu, Zhenghao and Li, Yishan and Yan, Yukun and Wang, Shuo and Liu, Zhiyuan and Gu, Yu and Yu, Minghe and Yu, Ge and Sun, Maosong},
year={2025}
url={https://arxiv.org/abs/2505.22095},
}
Contact Us
If you have questions, suggestions, or bug reports, please email us. We will try our best to help you.
hm.cypeng@gmail.com