# Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection

**Authors:** Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, Hannaneh Hajishirzi  
**arXiv:** [2310.11511](https://arxiv.org/abs/2310.11511)  
**Date:** 2023-10-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.11511v1/

## Key points

- Large language models (LLMs) often generate responses with factual inaccuracies due to heavy reliance on training data.
- Retrieval-Augmented Generation (RAG) supplements LLMs with external knowledge to improve response accuracy.
- Traditional RAG approaches indiscriminately retrieve fixed passages, limiting versatility and leading to unhelpful responses.
- Self-Reflective Retrieval-Augmented Generation (Self-RAG) empowers LLMs to adaptively retrieve passages and engage in self-reflection during generation.
- Self-RAG uses reflection tokens for the LLM to control behavior during inference, tailoring responses based on task requirements.
- Experimental evaluations show that Self-RAG outperforms existing models in fact verification, multiple-choice reasoning, open-domain question answering, biography writing, and long-form QA tasks.

## Summary

In the realm of large language models (LLMs), there is a common issue where responses generated by these models often contain factual inaccuracies due to their heavy reliance on the knowledge they have been trained on. To address this problem, researchers have developed Retrieval-Augmented Generation (RAG), a method that supplements LLMs with relevant external knowledge to improve response accuracy. However, traditional RAG approaches indiscriminately retrieve and incorporate a fixed number of passages, regardless of their relevance or necessity, which can limit the versatility of the LLM or lead to unhelpful responses. To enhance the quality and factuality of LLM-generated responses, a new framework called Self-Reflective Retrieval-Augmented Generation (Self-RAG) has been introduced. This innovative framework empowers an LLM to adaptively retrieve passages on-demand and engage in self-reflection during the generation process. By utilizing special tokens known as reflection tokens, the LLM can control its behavior during inference, allowing it to tailor its responses based on specific task requirements. Experimental evaluations of Self-RAG against state-of-the-art LLMs and retrieval-augmented models across various tasks demonstrate significant improvements in performance. In closed-set tasks such as fact verification and multiple-choice reasoning datasets, Self-RAG showcases superior accuracy compared to existing models. For short-form generation tasks like open-domain question answering, Self-RAG excels at providing accurate answers to diverse queries. Additionally, in long-form generation tasks such as biography writing and long-form QA, Self-RAG outperforms other models in terms of correctness, fluency, and citation precision. Overall, proves to be a groundbreaking framework that not only enhances the factuality and quality of LLM-generated responses but also improves their overall performance across a wide range of tasks.