Most RAG systems retrieve passages for every incoming query, even simple ones like 'what is 12 times 4' that the model could answer correctly on its own without any retrieved context, and even in cases where nothing in the index is actually relevant. Self-RAG, introduced by Asai et al. (2023), trains a model to address this by generating special reflection tokens during its own output. What is the specific problem these reflection tokens let the model address?
- They let the model retrieve passages in a language other than the one the query was written in, so the retriever can search a multilingual index without a separate translation step
- They let the model compress every retrieved passage into a shorter summary before generation, reducing the total number of tokens that must fit inside the context window
- They let the model decide on its own, per query, whether retrieval is even needed at all, and separately critique whether a retrieved passage is relevant and whether its own generated output is actually supported by that passage, rather than always retrieving and always trusting whatever was retrieved
- They let the model automatically retrain its own retriever component using the current query as a new labeled training example, improving retrieval quality over time
Why C? And why not the others?
Correct answer: C. They let the model decide on its own, per query, whether retrieval is even needed at all, and separately critique whether a retrieved passage is relevant and whether its own generated output is actually supported by that passage, rather than always retrieving and always trusting whatever was retrieved
Self-RAG's reflection tokens are produced by the model itself as part of ordinary generation and serve two purposes that always-retrieve pipelines lack: a retrieval-decision signal that lets the model judge, per query, whether retrieving is even necessary versus answering directly, and critique signals that assess whether a given retrieved passage is actually relevant and whether the model's own generated statements are properly supported by it; together these let the system skip retrieval on queries like simple arithmetic where it adds nothing, and flag or discount passages and claims that don't hold up, rather than blindly retrieving for every query and blindly trusting whatever came back. The reflection tokens have nothing to do with cross-lingual retrieval or translation; Self-RAG's mechanism is about deciding whether and how to use retrieval, not about which language is searched. They also are not a summarization mechanism; nothing in Self-RAG's design compresses retrieved passages into shorter text before generation. And the reflection tokens do not retrain the retriever at query time; Self-RAG's critic and reflection-token behavior are learned once during training on a fixed dataset, not updated online from each new incoming query.
Source: Asai, Wu, Wang, Sil & Hajishirzi, 'Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection' (2023), arXiv:2310.11511