Origin of Hallucination in LLMs, The physical source of hallucinations has found (arxiv.org)

A new study has identified specific neurons responsible for hallucinations in large language models (LLMs), addressing a critical issue where these models generate plausible yet factually incorrect outputs. Researchers discovered that less than 0.1% of the neurons in LLMs, termed H-Neurons, can predict the occurrence of hallucinations with remarkable accuracy. This discovery links the microscopic behavior of these neurons to macroscopic model performance, revealing that they contribute to over-compliance behaviors within the models. The significance of this research lies in its potential to enhance the reliability of LLMs during deployment. Understanding the origins and functions of H-Neurons, which emerge during pre-training, opens new avenues for refining training processes to mitigate hallucinations. By bridging behavioral patterns with underlying neural mechanisms, this work provides a foundation for future innovations aimed at creating more dependable AI systems and improves our understanding of how specific components of neural architectures can influence model outputs.