Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ReaRAG: Knowledge-guided Reasoning Enhances Factuality of Large Reasoning Models with Iterative Retrieval Augmented Generation

Zhicheng Lee, Shulin Cao, Jinxin Liu, Jiajie Zhang, Weichuan Liu, Xiaoyin Che, Lei Hou, Juanzi Li

TL;DR

ReaRAG introduces a knowledge-guided reasoning framework that grounds large reasoning models in external knowledge through iterative retrieval with a bounded reasoning chain. By distilling reasoning capabilities into a dedicated dataset and training under a Thought-Action-Observation paradigm, ReaRAG achieves strong factuality without RL-based policy optimization. The approach demonstrates competitive or superior performance on six QA benchmarks, particularly on harder multi-hop tasks like FRAMES and FanOutQA, while offering efficiency advantages over RL methods. The work highlights the value of reflection and error-correction in grounding reasoning trajectories, though it acknowledges limitations in action space, data construction efficiency, and inference latency. Overall, ReaRAG provides a practical, scalable alternative for improving factuality in retrieval-augmented reasoning while maintaining robust performance.

Abstract

Large Reasoning Models (LRMs) exhibit remarkable reasoning abilities but rely primarily on parametric knowledge, limiting factual accuracy. While recent works equip reinforcement learning (RL)-based LRMs with retrieval capabilities, they suffer from overthinking and lack robustness in reasoning, reducing their effectiveness in question answering (QA) tasks. To address this, we propose ReaRAG, a factuality-enhanced reasoning model that explores diverse queries without excessive iterations. Our solution includes a novel data construction framework with an upper bound on the reasoning chain length. Specifically, we first leverage an LRM to generate deliberate thinking, then select an action from a predefined action space (Search and Finish). For Search action, a query is executed against the RAG engine, where the result is returned as observation to guide reasoning steps later. This process iterates until a Finish action is chosen. Benefiting from ReaRAG's strong reasoning capabilities, our approach outperforms existing baselines on multi-hop QA. Further analysis highlights its strong reflective ability to recognize errors and refine its reasoning trajectory. Our study enhances LRMs' factuality while effectively integrating robust reasoning for Retrieval-Augmented Generation (RAG).

ReaRAG: Knowledge-guided Reasoning Enhances Factuality of Large Reasoning Models with Iterative Retrieval Augmented Generation

TL;DR

ReaRAG introduces a knowledge-guided reasoning framework that grounds large reasoning models in external knowledge through iterative retrieval with a bounded reasoning chain. By distilling reasoning capabilities into a dedicated dataset and training under a Thought-Action-Observation paradigm, ReaRAG achieves strong factuality without RL-based policy optimization. The approach demonstrates competitive or superior performance on six QA benchmarks, particularly on harder multi-hop tasks like FRAMES and FanOutQA, while offering efficiency advantages over RL methods. The work highlights the value of reflection and error-correction in grounding reasoning trajectories, though it acknowledges limitations in action space, data construction efficiency, and inference latency. Overall, ReaRAG provides a practical, scalable alternative for improving factuality in retrieval-augmented reasoning while maintaining robust performance.

Abstract

Large Reasoning Models (LRMs) exhibit remarkable reasoning abilities but rely primarily on parametric knowledge, limiting factual accuracy. While recent works equip reinforcement learning (RL)-based LRMs with retrieval capabilities, they suffer from overthinking and lack robustness in reasoning, reducing their effectiveness in question answering (QA) tasks. To address this, we propose ReaRAG, a factuality-enhanced reasoning model that explores diverse queries without excessive iterations. Our solution includes a novel data construction framework with an upper bound on the reasoning chain length. Specifically, we first leverage an LRM to generate deliberate thinking, then select an action from a predefined action space (Search and Finish). For Search action, a query is executed against the RAG engine, where the result is returned as observation to guide reasoning steps later. This process iterates until a Finish action is chosen. Benefiting from ReaRAG's strong reasoning capabilities, our approach outperforms existing baselines on multi-hop QA. Further analysis highlights its strong reflective ability to recognize errors and refine its reasoning trajectory. Our study enhances LRMs' factuality while effectively integrating robust reasoning for Retrieval-Augmented Generation (RAG).

Paper Structure

This paper contains 40 sections, 2 equations, 3 figures, 11 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Reasoning-enhanced LLMs.
  4. Retrieval-Augmented Generation.
  5. Reasoning-enhanced RAG.
  6. Methodology
  7. Task formulation
  8. Knowledge-guided reasoning chain generation
  9. Data construction.
  10. Data filtering.
  11. Factuality-enhanced LRMs: ReaRAG
  12. Fine-tuning.
  13. Inference.
  14. Experiments
  15. Experimental setup
  16. ...and 25 more sections

Figures (3)

  • Figure 1: Unlike LRMs, ReaRAG iteratively constructs knowledge-guided reasoning chains for factual answers.
  • Figure 2: Overview of our approach to develop a factuality-enhanced reasoning model ReaRAG. To equip ReaRAG with knowledge-guided reasoning ability, we propose an automated data construction approach (Algorithm \ref{['alg:data_construction']}). Next, we fine-tune ReaRAG on the constructed dataset to conduct reasoning iteratively, following the Thought-Action-Observation Paradigm to solve complex queries. Pseudocode for the inference stage is provided in Algorithm \ref{['alg:inference']}.
  • Figure 3: Comparison of chain length across all benchmarks. We measure the number of retrievals needed for the baselines to achieve a full $\text{ACC}_{L}$ score. Search-o1 consistently requires more steps than ReaRAG, highlighting the tendency of the underlying LRM to overthink in multi-hop QA tasks.