pat-jj commited on
Commit
2a38f91
·
verified ·
1 Parent(s): ce47dc6

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +24 -0
README.md CHANGED
@@ -29,4 +29,28 @@ Here's the user query:
29
  <|im_start|>assistant
30
  Let me write the SQL query with reasoning.
31
  <think>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
32
  ```
 
29
  <|im_start|>assistant
30
  Let me write the SQL query with reasoning.
31
  <think>
32
+ ```
33
+
34
+
35
+ # DeepRetrieval
36
+ ## Overview
37
+
38
+ DeepRetrieval is a novel approach that uses reinforcement learning (RL) to train Large Language Models (LLMs) for query generation without requiring supervised data. Instead of relying on expensive human-annotated or distilled reference queries, DeepRetrieval enables LLMs to learn through direct trial and error, using retrieval metrics as rewards.
39
+ ## Key Features
40
+
41
+ - **No Supervision Required**: Eliminates the need for expensive human-annotated or distilled reference queries
42
+ - **RL-Based Framework**: Uses reinforcement learning to optimize query generation directly for retrieval performance
43
+ - **State-of-the-Art Performance**: Achieves remarkable results across diverse retrieval tasks
44
+
45
+ Please view our [GitHub page](https://github.com/pat-jj/DeepRetrieval) for instructions.
46
+
47
+ [DeepRetrieval Paper](arxiv.org/abs/2503.00223)
48
+ ```
49
+ @article{jiang2025deepretrievalhackingrealsearch,
50
+ title={DeepRetrieval: Hacking Real Search Engines and Retrievers with Large Language Models via Reinforcement Learning},
51
+ author={Pengcheng Jiang and Jiacheng Lin and Lang Cao and Runchu Tian and SeongKu Kang and Zifeng Wang and Jimeng Sun and Jiawei Han},
52
+ year={2025},
53
+ journal = {arXiv preprint arXiv: 2503.00223},
54
+ url={https://arxiv.org/abs/2503.00223}
55
+ }
56
  ```