Using LLM for confidence scoring is a bad idea

The article from rumple.github.io discusses the limitations of using Large Language Models (LLMs) for confidence scoring. LLMs are used to evaluate the confidence of their own responses, but this approach has its drawbacks. The author argues that LLMs are not reliable for confidence scoring due to their inherent biases and lack of understanding of the context. This can lead to overconfidence in incorrect responses and underconfidence in correct ones. The author suggests that human evaluation and domain-specific knowledge are more effective for confidence scoring.