# RAG+: Enhancing Retrieval-Augmented Generation with Application-Aware Reasoning

**Authors:** Yu Wang, Shiwan Zhao, Ming Fan, Zhihu Wang, Yubo Zhang, Xicheng Zhang, Zhengfan Wang, Heyuan Huang, Ting Liu  
**arXiv:** [2506.11555](https://arxiv.org/abs/2506.11555)  
**Date:** 2025-06-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2506.11555v1/

## Key points

- Authors introduce RAG+ as an extension to the existing Retrieval-Augmented Generation (RAG) framework for integrating external knowledge into large language models (LLMs)
- RAG+ incorporates application-aware reasoning by introducing a dual corpus of knowledge data and aligned application examples
- During inference, RAG+ retrieves information from both knowledge corpus and application examples simultaneously
- Experimental results show that RAG+ consistently outperforms standard RAG variants across mathematical, legal, and medical domains
- On average, RAG+ achieves improvements ranging from 3% to 5%, with peak gains of up to 7.5% in complex scenarios

## Summary

, , , , 

In their paper titled "RAG+: Enhancing Retrieval-Augmented Generation with Application-Aware Reasoning," authors Yu Wang, Shiwan Zhao, Ming Fan, Zhihu Wang, Yubo Zhang, Xicheng Zhang, Zhengfan Wang, Heyuan Huang, and Ting Liu introduce a novel approach to integrating external knowledge into large language models (LLMs) for knowledge-intensive tasks. The existing framework of Retrieval-Augmented Generation (RAG) has proven effective in enhancing LLMs by retrieving relevant information from external sources. However, traditional RAG paradigms often fall short in bridging the gap between retrieved facts and task-specific reasoning processes. To address this limitation, the authors propose RAG+, an innovative extension that incorporates application-aware reasoning into the RAG pipeline. RAG+ introduces a dual corpus comprising both knowledge data and aligned application examples. These examples can be curated manually or generated automatically to ensure alignment with specific tasks. During inference, RAG+ retrieves information from both the knowledge corpus and the application examples simultaneously. This design enables LLMs not only to access pertinent information but also to apply it within structured and goal-oriented reasoning processes. Through experiments conducted across mathematical, legal, and medical domains using multiple models, the authors demonstrate that RAG+ consistently outperforms standard RAG variants. On average, RAG+ achieves improvements ranging from 3% to 5%, with peak gains of up to 7.5% in complex scenarios. By integrating retrieval with actionable application through a more cognitively grounded framework for knowledge integration, RAG+ represents a significant step towards developing more interpretable and capable LLMs. This research contributes valuable insights into advancing the effectiveness of large language models for various knowledge-intensive tasks across different domains.