AI RESEARCH

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

arXiv CS.AI

ArXi:2604.04743v1 Announce Type: cross Large language models (LLMs) hallucinate: they produce fluent outputs that are factually incorrect. We present a geometric dynamical systems framework in which hallucinations arise from task-dependent basin structure in latent space.