File size: 1,387 Bytes
881a485
 
 
4d90a39
 
 
 
 
881a485
 
4d90a39
 
881a485
 
4d90a39
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
---
base_model:
- intfloat/e5-base-v2
language:
- en
license: mit
pipeline_tag: text-retrieval
library_name: transformers
---

# Agentic-R: Learning to Retrieve for Agentic Search

## Introduction

This is the **Agentic-R** retriever model introduced in the paper [Agentic-R: Learning to Retrieve for Agentic Search](https://huggingface.co/papers/2601.11888).

Agentic-R is a dense retriever specifically tailored for agentic search, where an agent interleaves multi-step reasoning with on-demand retrieval. Unlike retrievers designed for single-turn retrieval-augmented generation (RAG) that rely only on local passage utility, Agentic-R uses both local query-passage relevance and global answer correctness to measure passage utility in a multi-turn agentic search.

For detailed usage instructions, training scripts, and evaluation code, please refer to the [🧩 GitHub repository](https://github.com/8421BCD/Agentic-R).

## Citation

If you find this work helpful, please cite our paper:

```bibtex
@misc{liu2026agenticrlearningretrieveagentic,
      title={Agentic-R: Learning to Retrieve for Agentic Search}, 
      author={Wenhan Liu and Xinyu Ma and Yutao Zhu and Yuchen Li and Daiting Shi and Dawei Yin and Zhicheng Dou},
      year={2026},
      eprint={2601.11888},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2601.11888}, 
}
```