Self-RAG paper

Definition: Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection is a paper by Akari Asai and colleagues introducing Self-Reflective Retrieval-Augmented Generation (Self-RAG). The framework trains a language model to decide when to retrieve external passages and to assess those passages and its own answers using special reflection tokens.

Key facts

• The arXiv record dates the preprint to 17 October 2023 and lists Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil and Hannaneh Hajishirzi as authors.
• As of 17 October 2026, the authors' implementation repository describes retrieval on demand: the model can retrieve repeatedly during generation or skip retrieval entirely.
• The paper's reflection-token table separates the retrieval decision from assessments of passage relevance, evidence support and response usefulness.

How does the model decide when to retrieve?

Retrieval-augmented generation combines retrieved text passages with language model generation. In Self-RAG's inference procedure, the model predicts a retrieval token from the input and the answer generated so far. The experiments treat a sentence as an output segment.

When retrieval is needed, an external retriever finds passages using the question and preceding generation. The model processes candidate passages in parallel, assesses their relevance and generates a continuation conditioned on each passage. It then evaluates the continuations and selects between them. When retrieval is unnecessary, it generates the next segment directly and assesses its usefulness.

The paper also permits a retrieval threshold: retrieval starts when the normalized probability of the affirmative retrieval token exceeds the selected threshold. This makes retrieval frequency adjustable during inference, the stage when a trained model produces an answer.

What does self-critique measure?

The paper's token definitions distinguish several judgments. IsRel assesses whether a passage contains useful information for the input. IsSup assesses whether the passage supports the statements in the generated segment, with values for full support, partial support or no support. IsUse rates the usefulness of the response to the input on a scale from 1 to 5.

The authors' repository quick start shows these special tokens in model prompts or outputs:

[Retrieval]
[Relevant]
[Fully supported]
[Utility:5]

These are learned model assessments. In the paper's inference procedure, they influence which continuation is selected. A passage can be relevant to the question while only partially supporting a particular answer, so relevance and support have separate token categories.

How is reflection learned and used to select an answer?

The training procedure first uses GPT-4 feedback to train a critic model, a model that predicts reflection labels. That critic inserts reflection tokens into training examples alongside retrieved passages. The generator then learns to predict both answer text and reflection tokens with a standard next-token prediction objective. It produces its own reflection tokens during inference without depending on the separate critic.

For answer selection, the paper uses segment-level beam search: it retains several candidate continuations while building an answer. Candidate scoring combines the generation probability with weighted critique scores derived from token probabilities. Increasing the weight assigned to evidence support gives that assessment more influence. The paper also describes a hard constraint that removes a continuation when its support token indicates no support.

Related terms

• Retrieval-augmented generation (RAG)
• Retrieval reranking
• Citation faithfulness
• Large language model (LLM)

FAQ

Does Self-RAG always search before answering?

The authors' project explanation describes retrieval decisions at each segment. The model can retrieve several times or complete an answer without retrieval.

Does self-critique prove that an answer is correct?

The paper's ethical concerns section acknowledges that the model can still produce outputs whose citations provide incomplete support. Reflection tokens are model predictions used to score and filter candidate continuations.

Sources

1. arXiv, "Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection", 17 October 2023
2. Akari Asai and colleagues, "Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection", preprint, 17 October 2023
3. Akari Asai and colleagues, "Self-RAG implementation repository", undated
4. Akari Asai and colleagues, "Self-RAG project page", undated

---

Check how a page reads to answer engines. Open the GEO score analyzer