Self-critique lets a model evaluate retrieved evidence and its own generated answer instead of treating retrieval and generation as passive, fixed steps. In Self-RAG, the model can retrieve passages when needed, skip retrieval when its parameters already contain the answer, and reflect during generation. section 9.0
Reflection tokens make the evaluations explicit and controllable at inference time. IsRel filters irrelevant passages, IsSup checks claims against evidence, and IsUse judges whether the resulting answer is worth returning. section 9.8
Self-critique supports two deployment patterns:
- Prompted Self-RAG emulates reflection tokens with separate grader calls, adding an API round-trip for each critique. section 9.8
- Trained Self-RAG fine-tunes a smaller model to emit reflection tokens inline, shifting the cost to training rather than repeated critique calls. section 9.8
- Decoding weights adjust strictness without retraining: higher pressure against unverified statements suits compliance work, while looser settings support creative work. section 9.8
Adaptive answer checking
A query whose answer is already stored in the model can skip search. A query requiring external evidence can retrieve passages, discard irrelevant ones, check claims against them, and reject an answer judged unworthy. The same mechanism can penalize unsupported statements more strongly for compliance than for creative generation. section 9.0 section 9.8
Self-critique is not independent verification: a model judging its own output may share the assumptions that produced it. External evaluators remain necessary when that blind spot matters. section 9.8
