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

Retrieval-Augmented Generation (RAG) lifts the factuality of Large Language Models (LLMs) by injecting external knowledge, yet it falls short on problems that demand multistep inference; conversely, purely reasoningoriented approaches often hallucinate or misground facts. This survey synthesizes both strands under a unified reasoning-retrieval perspective. We first map how advanced reasoning optimizes each stage of RAG (Reasoning- Enhanced RAG). Then, we show how retrieved knowledge of different type supply missing premises and expand context for complex inference (RAG-Enhanced Reasoning). Finally, we spotlight emerging Synergized RAG-Reasoning frameworks, where (agentic) LLMs iteratively interleave search and reasoning to achieve state-of-the-art performance across knowledge-intensive benchmarks. We categorize methods, datasets, and open challenges, and outline research avenues toward deeper RAG-Reasoning systems that are more effective, multimodally-adaptive, trustworthy, and human-centric. The collection is available at https://github.com/ DavidZWZ/Awesome-RAG-Reasoning.

Introduction. The remarkable progress in Large Language Models (LLMs) has transformed a wide array of fields, showcasing unprecedented capabilities across diverse tasks (Zhao et al., 2023). Despite these advancements, the effectiveness of LLMs remains hindered by two fundamental limitations: knowledge hallucinations, due to the static and parametric manner of their knowledge storage (Huang et al., 2025b); and struggles with complex reasoning, especially when tackling real-world problems (Chang et al., 2024). These limitations have driven the development of two major directions: Retrieval- Augmented Generation (RAG) (Fan et al., 2024a), which provides LLMs with external knowledge; and various methods aimed at enhancing their inherent reasoning abilities (Chen et al., 2025c). The two limitations are inherently intertwined: missing knowledge can impede reasoning, and flawed reasoning hinders knowledge utilization (Tonmoy et al., 2024). Naturally, researchers have increasingly explored combining retrieval with reasoning, though early work followed two separate, one-way enhancements.

Discussion / Conclusion. Future research directions for Synergized RAG- Reasoning systems center around enhancing both reasoning and retrieval capabilities to meet realworld demands for accuracy, efficiency, trust, and user alignment. We outline several key challenges and opportunities below. This survey charts the rapid convergence of retrieval and reasoning in LLMs. We reviewed three evolutionary stages: (1) Reasoning-Enhanced RAG, which uses multi-step reasoning to refine each stage of RAG; (2) RAG-Enhanced Reasoning, which leverages retrieved knowledge to bridge factual gaps during long CoT; and (3) Synergized RAG-Reasoning systems, where single- or multiagents iteratively refine both search and reasoning, exemplified by recent “Deep Research” platforms. Collectively, these lines demonstrate that tight retrieval–reasoning coupling improves factual grounding, logical coherence, and adaptability beyond one-way enhancement. Looking forward, we identify research avenues toward synergized RAG-Reasoning systems that are more effective, multimodally-adaptive, trustworthy, and humancentric.