# Towards Agentic RAG with Deep Reasoning: A Survey of RAG-Reasoning Systems in LLMs

**Authors:** Yangning Li, Weizhi Zhang, Yuyao Yang, Wei-Chieh Huang, Yaozu Wu, Junyu Luo, Yuanchen Bei, Henry Peng Zou, Xiao Luo, Yusheng Zhao, Chunkit Chan, Yankai Chen, Zhongfen Deng, Yinghui Li, Hai-Tao Zheng, Dongyuan Li, Renhe Jiang, Ming Zhang, Yangqiu Song, Philip S. Yu  
**arXiv:** [2507.09477](https://arxiv.org/abs/2507.09477)  
**Date:** 2025-07-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2507.09477v1/

## Key points

- Comprehensive survey on Retrieval-Augmented Generation (RAG) and reasoning with Large Language Models (LLMs)
- Synthesis of over 200 research papers to create a unified taxonomy for advanced reasoning techniques in RAG
- Categorization into three main frameworks: Reasoning-Enhanced RAG, RAG-Enhanced Reasoning, and Synergized RAG-Reasoning systems
- Emphasis on optimizing each stage of RAG through multi-step reasoning and leveraging retrieved knowledge for complex inference tasks
- Future research focus on genuine multimodality with Multi-modal Large Language Models (MLLMs) and agentic capabilities through hybrid-modal chain-of-thought reasoning
- Importance of retrieval trustworthiness for reliable downstream reasoning in Synergized RAG-Reasoning systems
- Suggestions for enhancing system traceability with techniques like watermarking and digital fingerprinting
- Need for dynamic and adaptive methods to combat adversarial attacks and ensure system robustness
- Tight coupling between retrieval and reasoning improves factual grounding, logical coherence, and adaptability in LLMs

## Summary

This comprehensive survey explores the intersection of Retrieval-Augmented Generation (RAG) and reasoning with Large Language Models (LLMs). The authors synthesize over 200 research papers to provide a unified taxonomy that encompasses advanced reasoning techniques in RAG as well as the integration of retrieved knowledge for complex inference tasks. The scope of the survey prioritizes breadth over depth and categorizes methods into three main frameworks: Reasoning-Enhanced RAG, RAG-Enhanced Reasoning, and Synergized RAG-Reasoning systems. These frameworks focus on optimizing each stage of RAG through multi-step reasoning, leveraging retrieved knowledge for complex inference tasks, and combining search and reasoning iteratively to achieve state-of-the-art performance across knowledge-intensive benchmarks. The authors also highlight the need for future research to move beyond traditional vision-text paradigms towards genuine multimodality by strengthening foundational abilities of Multi-modal Large Language Models (MLLMs) such as grounding and cross-modal reasoning. They also emphasize enhancing agentic capabilities through hybrid-modal chain-of-thought reasoning for real-world interaction via multimodal search tools. Furthermore, retrieval trustworthiness is crucial in maintaining reliable downstream reasoning in Synergized RAG-Reasoning systems. Techniques like watermarking and digital fingerprinting are suggested to enhance system traceability. Future research should focus on developing dynamic and adaptive methods to combat adversarial attacks and ensure system robustness. In conclusion, this survey charts the rapid convergence of retrieval and reasoning in LLMs, showcasing how tight coupling between retrieval and reasoning improves factual grounding, logical coherence, and adaptability. The authors identify research avenues towards deeper RAG-Reasoning systems that are more effective, multimodally-adaptive, trustworthy, and human-centric. The collection of resources related to this survey can be found at https://github.com/DavidZWZ/Awesome-RAG-Reasoning.