pat-jj commited on
Commit
d24c768
·
verified ·
1 Parent(s): 13711d7

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +23 -0
README.md CHANGED
@@ -29,4 +29,27 @@ 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
+ # DeepRetrieval
35
+ ## Overview
36
+
37
+ 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.
38
+ ## Key Features
39
+
40
+ - **No Supervision Required**: Eliminates the need for expensive human-annotated or distilled reference queries
41
+ - **RL-Based Framework**: Uses reinforcement learning to optimize query generation directly for retrieval performance
42
+ - **State-of-the-Art Performance**: Achieves remarkable results across diverse retrieval tasks
43
+
44
+ Please view our [GitHub page](https://github.com/pat-jj/DeepRetrieval) for instructions.
45
+
46
+ [DeepRetrieval Paper](arxiv.org/abs/2503.00223)
47
+ ```
48
+ @article{jiang2025deepretrievalhackingrealsearch,
49
+ title={DeepRetrieval: Hacking Real Search Engines and Retrievers with Large Language Models via Reinforcement Learning},
50
+ author={Pengcheng Jiang and Jiacheng Lin and Lang Cao and Runchu Tian and SeongKu Kang and Zifeng Wang and Jimeng Sun and Jiawei Han},
51
+ year={2025},
52
+ journal = {arXiv preprint arXiv: 2503.00223},
53
+ url={https://arxiv.org/abs/2503.00223}
54
+ }
55
  ```