“The Self-RAG framework trains a language model to reflect on its own generation by outputting special tokens that decide whether to retrieve documents and critique the generated text.”