Tag
This paper introduces diversion decoding, a novel method for detecting hallucinations in large language models by actively challenging model responses during the decoding phase to extract features for training an uncertainty heuristic, achieving superior performance with lower computational complexity.