Georgia Institute of Technology College of Sciences

Large language models often guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty. Such "hallucinations" persist even in state-of-the-art systems. We analyze this phenomenon from a mathematical perspective and find that the statistical pressures of current training pipelines induce hallucinations; moreover, current evaluation procedures reward guessing over acknowledging uncertainty. The talk will be fact-based, and the speaker will readily admit ignorance.