Hallucination Basins: A Dynamic Framework for Understanding and Controlling LLM Hallucinations

Kalyan Cherukuri, Lav R. Varshney

Large language models (LLMs) hallucinate: they produce fluent outputs that are factually incorrect. We present a geometric dynamical systems framework in which hallucinations arise…