From Architecture to Output: Structural Origins of Hallucination in Large Language Models and the Amplifying Role of Data

Md. Rejaul Korim Sadi, Toufiqur Rahman Tasin, Golam Mostofa Naeem

Large language models produce fluent, confident, factually wrong output. Existing taxonomies classify these failures by output type -- intrinsic versus extrinsic, faithfulness vers…