.png){kind=icon}
.png){kind=icon}
Return
ReflectRAG: Enhancing retrieval-augmented generation with GRPO-optimized iterative reflection
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2026.134047.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Retrieval-Augmented Generation (RAG) enhances question answering (QA) by grounding responses in external documents, yet struggles with irrelevant retrieval and single-pass generation errors. To address these issues, we propose ReflectRAG, a framework that advances RAG through structured planning, iterative reflection, and reinforcement learning (RL). Specifically, ReflectRAG first retrieves and re-ranks documents and generates answers guided by fact-based plans. Then, ReflectRAG leverages a reflection module to iteratively critique and refine plans and answers, and optimizes the reflection module via RL. By integrating planning, reflection, and RL, ReflectRAG ensures precise, contextually aligned answers, offering a robust solution for open-domain QA. To demonstrate the effectiveness of our proposed model, we carry out extensive experiments on MS MARCO v2.1, TriviaQA (TQA), and Natural Questions (NQ). The comparative results on three open-domain datasets show that ReflectRAG achieves up to +0.31 absolute F1 points over the BASE RAG baseline across MS MARCO v2.1, TriviaQA, and NQ, establishing ReflectRAG as an effective solution for enhancing factual accuracy and coherence in QA systems. The source code is available at https://github.com/uncxlirh/ReflectRAG .
Keywords:
Retrieval-Augmented Generation
Iterative Reflection
Reinforcement Learning
Question Answering
Structured Planning
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W