44 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Why Pretrained Models Hallucinate
Georgia Institute of Technology
Recording
Abstract
This talk analyzes hallucination in language models from a mathematical perspective and shows that the statistical pressures of the next-word-prediction training pipeline induce hallucinations under fairly general conditions, even when models are trained on clean data. The core proof is a simple reduction from a natural classification task, deciding whether a statement is a fact or a hallucination, to a generation or prompt-completion task.