A language model answering a trivia question may sound confident even when it does not know the answer. Conformal factuality is a statistical method for reducing such factual errors, often called hallucinations. section 5.7
The method applies a predefined scoring function to claims in a response and filters out claims the model is less confident about. The remaining response can be held to a selected probability standard—for example, at least a 90% chance that all its claims are correct. section 5.7
That assurance comes at a cost: filtering may remove a significant part of the original answer. The method is described as impractical today, and it reduces rather than solves hallucinations. section 5.7
