Using LLM for confidence scoring is a bad idea
rumple.github.iojimauthors
An analysis examining why large language models are unsuitable for generating confidence scores, arguing against using LLM outputs as reliability metrics for predictions or classifications.
Why it mattersTool builders relying on LLM confidence signals may deploy systems with inflated or misaligned reliability estimates, potentially leading to false trust in model outputs.
Read at rumple.github.io