Public demo

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

Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, Hannaneh Hajishirzi (2023)

Self-RAG trains a model to decide when to retrieve, how to use retrieved passages, and how to critique its own generated output with reflection tokens.

Why it matters

Self-RAG moves beyond always-on retrieval. It frames retrieval as a controlled behavior that can be triggered, judged, and refined, which is close to the writing workflow where researchers ask for evidence only where a claim needs support.

Highlights

  • [method] Self-RAG introduces reflection tokens for retrieval decisions, relevance, support, and utility. — why: these tokens make retrieval behavior more explicit and controllable.
  • [claim] The model can adaptively retrieve instead of retrieving for every generation step. — why: this matters for efficiency and for avoiding unnecessary context.
  • [finding] Self-reflection improves factuality and task performance compared with standard generation and retrieval baselines. — why: it shows that critique signals can improve grounded generation.
  • [limitation] Reflection is still model-generated and must not be treated as external verification. — why: this boundary mirrors SAV's "verify before citing" principle.

Reader notes

  • Pair this with retrieval-augmented-generation: RAG adds documents; Self-RAG adds a control loop around when and how documents are used.
  • Good demo example for showing why evidence support should be visible beside the text, not hidden inside the model.
  • Reflection tokens are conceptually useful for product design: retrieval, support, and utility can become user-facing states.

Extracted evidence examples

  • OBJECTIVE: This work introduces Self-Reflective Retrieval-augmented Generation (SELF-RAG) to improve an LLM's generation quality, including its factual accuracy without hurting its versatility, via on-demand retrieval and self-reflection.
    This work introduces Self-Reflective Retrieval-augmented Generation (SELF-RAG) to improve an LLM's generation quality, including its factual accuracy without hurting its versatility, via on-demand retrieval and self-reflection.
  • OBJECTIVE: Approaches are described to enforce control during the inference process to meet distinct task objectives.
    In this section, we describe approaches to enforce control to meet these distinct objectives during the inference process.
  • RESEARCH_PROBLEM: Despite scaling, LLMs still produce factual errors, necessitating improvements to prevent real-world harms like misinformation and dangerous advice.
    State-of-the-art LLMs continue to struggle with factual errors (Mallen et al., 2023;Min et al., 2023) despite their increased model and data scale (Ouyang et al., 2022).
    This work aims to improve the factuality of LLM outputs, the lack of which continues to cause numerous real-world problems (e.g., spread of misinformation and provision of incorrect and dangerous advice).
  • RESEARCH_PROBLEM: Managers must delegate tasks to their team to successfully complete the extensive list required for event success.
    Input Give an example of a situation where you have to delegate tasks (from gpt4 alpaca) Output [ Retrieve =No]The manager realizes that there is an extensive list of tasks that need to be completed in order to make the event a success.[
    ISSUP =Partially Supported][ Retrieve =No]Since it is not possible for the manager to handle everything alone, they must delegate tasks to their team members.[
  • RESEARCH_PROBLEM: Retrieval-Augmented Generation (RAG) methods improve knowledge-intensive tasks by augmenting LM input space with retrieved text passages.
    Retrieval-Augmented Generation (RAG) augments the input space of LMs with retrieved text passages (Guu et al., 2020;Lewis et al., 2020), leading to large improvements in knowledge-intensive tasks after fine-tuning or used with off-the-shelf LMs (Ram et al., 2023).
  • RESEARCH_PROBLEM: German Shepherd -a simple difference of opinion, it is the fundamental issue of the breedś essential conformation and movement.
    Retrieve =Yes]<p>German Shepherd -a simple difference of opinion, it is the fundamental issue of the breedś essential conformation and movement.
  • RESEARCH_PROBLEM: Word embedding is a set of language modeling and feature learning techniques in NLP where words or phrases are mapped to vectors of real numbers.
    Evidence Word embedding Word embedding is the collective name for a set of language modeling and feature learning techniques in natural language processing (NLP) where words or phrases from the vocabulary are mapped to vectors of real numbers.
  • RESEARCH_GAP: Existing RAG methods indiscriminately retrieve passages, failing to guarantee factual support or fine-grained self-reflection, hindering LLM versatility and leading to low-quality, inconsistent generations.
    However, these methods may hinder the versatility of LLMs or introduce unnecessary or off-topic passages that lead to low-quality generations (Shi et al., 2023) since they retrieve passages indiscriminately regardless of whether the factual grounding is helpful.
    Moreover, the output is not guaranteed to be consistent with retrieved relevant passages (Gao et al., 2023) since the models are not explicitly trained to leverage and follow facts from provided passages.