← All questions
AgentsMedium· Self-Reflection· Training & Fine-tuning· Evaluation

Self-RAG’s critic is the generator itself. Isn’t that j...

It is important to separate the timelines. At training time, the judgment is external—a GPT-4 teacher distilled through a critic model into corpus annotations. At serving time, critique and content share a forward pass, so errors can correlate. Self-RAG works because the judgment was learned from outside, providing a genuine check and balance rather than relying on the model’s own unguided reflection.

Code playground (Python)