Comment by Jiro

From that abstract it doesn't sound like they allowed for the possibility that the LLM could be trained to say "I don't know" for some things.