We introduce BGE-Reasoner, an end-to-end reasoning-intensive information retrieval framework. BGE-Reasoner is characterized by three key features:

1. End-to-end: It comprises three core components in IR—BGE-Reasoner-Rewriter, BGE-Reasoner-Embed, and BGE-Reasoner-Reranker—covering the entire retrieval pipeline, from query rewriting and retrieval to reranking for reasoning-intensive tasks.
2. Excellent performance: BGE-Reasoner achieves state-of-the-art (SOTA) performance on BRIGHT, a reasoning-intensive information retrieval benchmark, with an nDCG@10 of 46.4 across 12 datasets (BGE-Reasoner-0928, released on Oct 11, 2025), outperforming the previous SOTA by +0.6 points (45.8 from DIVER, Aug 27, 2025).
3. Open-source resources: We have fully released the related resources to facilitate future research on reasoning-intensive information retrieval.

For the training details of BGE-Reasoner-Embed, please refer to this page: ReasonEmbed.

For the training details of BGE-Reasoner-Reranker, please refer to this page: Retro-star.

1. Query Rewrite: BGE-Reasoner-Rewriter generates 5 rewritten queries for each original query; all 5 rewrites are used for retrieval.
2. Retrieval: For each rewritten query, BGE-Reasoner-Embed and BM25 retrieve the top-2000 documents. We aggregate results across the 5 rewrites by summing the corresponding scores to produce a final score per method.
3. Reranking:
   • We rerank the top-100 documents from each retrieval method using BGE-Reasoner-Reranker (models: 8B, 14B, 32B), producing 6 reranked top-10 lists (2 retrieval methods × 3 reranker sizes).
   • We also create a hybrid top-10 by fusing BGE-Reasoner-Embed and BM25 (weights: 0.75 / 0.25 after min–max normalization).
   • Finally, we combine the 7 top-10 lists (6 reranked + 1 hybrid) to produce the final top-10.

Pipeline:

Results:

Note:

• "Avg - ALL" refers to the average performance across all 12 datasets in the BRIGHT benchmark.
• "Avg - SE" refers to the average performance across the 7 datasets in the StackExchange subset of the BRIGHT benchmark.
• "Avg - CD" refers to the average performance across the 2 datasets in the Coding subset of the BRIGHT benchmark.
• "Avg - MT" refers to the average performance across the 3 datasets in the Theorem-based subset of the BRIGHT benchmark.

Sources of results:

[1] https://arxiv.org/pdf/2504.20595

[2] https://github.com/Debrup-61/RaDeR

[3] https://huggingface.co/ielabgroup/Rank-R1-32B-v0.2

[4] https://github.com/jataware/XRR2

[5] http://arxiv.org/pdf/2508.07050

[6] https://arxiv.org/pdf/2508.07995

Pipeline:

Results:

Note:

• "Avg - ALL" refers to the average performance across all 12 datasets in the BRIGHT benchmark.
• "Avg - SE" refers to the average performance across the 7 datasets in the StackExchange subset of the BRIGHT benchmark.
• "Avg - CD" refers to the average performance across the 2 datasets in the Coding subset of the BRIGHT benchmark.
• "Avg - MT" refers to the average performance across the 3 datasets in the Theorem-based subset of the BRIGHT benchmark.

Sources of results:

[1] https://arxiv.org/pdf/2504.20595

[2] https://github.com/Debrup-61/RaDeR

[3] https://huggingface.co/ielabgroup/Rank-R1-32B-v0.2

[4] https://github.com/jataware/XRR2

[5] http://arxiv.org/pdf/2508.07050

[6] https://arxiv.org/pdf/2508.07995

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}
}

@article{lan2025retro,
  title={Retro*: Optimizing LLMs for Reasoning-Intensive Document Retrieval},
  author={Lan, Junwei and Chen, Jianlyu and Liu, Zheng and Li, Chaofan and Bao, Siqi and Lian, Defu},
  journal={arXiv preprint arXiv:2509.24869},
  year={2025}
}