The result seems to rely on stipulating the LLM must answer true or false to all its questions, and can't say "I don't know." So it's an interesting result, but

The result seems to rely on stipulating the LLM must answer true or false to all its questions, and can't say "I don't know." So it's an interesting result, but it's not obvious that it tells us much about our actual problem, which is 100% about how to get a system that accurately understand the level of confidence it should have.