LimRank: Less is More for Reasoning-Intensive Information Reranking

This repository contains the limrank-7b model, based on Qwen2.5-7B, which was presented in the paper LimRank: Less is More for Reasoning-Intensive Information Reranking.

LimRank demonstrates an efficient approach to adapt modern Large Language Models (LLMs) for reasoning-intensive information reranking tasks. This is achieved by leveraging LIMRANK-SYNTHESIZER, a reusable and open-source pipeline that generates minimal yet high-quality synthetic supervision data. Through this approach, LimRank achieves competitive performance on challenging benchmarks like BRIGHT and FollowIR, utilizing less than 5% of the data typically required by prior methods. The model also shows strong generalization capabilities across various downstream tasks, including scientific literature search and retrieval-augmented generation.

Links

• Paper: LimRank: Less is More for Reasoning-Intensive Information Reranking
• Code/GitHub Repository: https://github.com/SighingSnow/LimRank
• Trained LimRank Model: songtingyu/limrank-7b
• Training Datasets: songtingyu/limrank-data
• Evaluation Results: sogntingyu/limrank-results
• Running Files for Reproduction: songtingyu/limrank-run-files

Citation

If you find our paper useful, please cite our work:

@misc{song2025limrankreasoningintensiveinformationreranking,
      title={LimRank: Less is More for Reasoning-Intensive Information Reranking}, 
      author={Tingyu Song and Yilun Zhao and Siyue Zhang and Chen Zhao and Arman Cohan},
      year={2025},
      eprint={2510.23544},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.23544}, 
}

Acknowledgements

We would like to thank the authors of the following papers and repos for their open-source contributions.

• rank1
• ReasonIR
• MTEB

License

The model is released under the MIT License.