We propose ReasonEmbed, a new text embedding model for reasoning-intensive document retrieval based on innovations of how synthetic data is generated and used. Our work includes the following technical contributions.

1. We design a novel data synthesis method, called ReMixer.
2. We introduce a self-adaptive training method tailored for our synthetic data, termed Redapter.
3. We implement ReasonEmbed based on multiple LLM backbones of varying model sizes, which achieve state-of-the-art (SOTA) performance on reasoning-intensive document retrieval tasks. Notably, our model built on Qwen3-4B reaches an nDCG@10 score of 37.1 on the BRIGHT benchmark, which already surpasses all existing text embedding models. While the Qwen3-8B based varient improves the performance to 38.1. Moreover, on the R2MED benchmark, ReasonEmbed-Qwen3-8B attains an nDCG@10 score of 43.18, surpassing all of the existing models by a large margin and leading to new SOTA performance.

This paper has been accepted to the ACL 2026 Main Conference (Oral Presentation)🎉. For more details, please refer to our paper.

Evaluation results of ReasonEmbed on the BRIGHT benchmark:

Evaluation results of ReasonEmbed on the R2MED benchmark:

If you find this repository useful, please consider giving a star ⭐ and citation:

@article{chen2025reasonembed,
  title={ReasonEmbed: Enhanced Text Embeddings for Reasoning-Intensive Document Retrieval},
  author={Chen, Jianlyu and Lan, Junwei and Li, Chaofan and Lian, Defu and Liu, Zheng},
  journal={arXiv preprint arXiv:2510.08252},
  year={2025}
}