The Half-Truth Prince: Hallucinations in Language Models

Imagination is more important than knowledge, for knowledge is limited…” Einstein might have had mortals in mind while stating this, but Large Language Models (LLMs), which have transformed the face of Artificial Intelligence (AI), have proved to be worthy of this quote as well because of a curious phenomenon: AI Hallucinations - where models fabricate information with unwarranted confidence. This paper provides a comprehensive investigation of hallucination, the types of hallucination LLMs will exhibit, and the causes behind it. Also, to experimentally investigate hallucination limits, A new strategy that utilizes complex, purposely dishonest multiple-choice questions based on Data Structures (trees and graphs) has been introduced. The results show that LLMs definitively select and justify incorrect answers, clearly demonstrating limitations as well as their creative reasoning ability across multiple domains. The paper also examines the current strategies to mitigate hallucinations, such as prompt engineering and fine-tuning temperature, in order to provide a more sophisticated argument to converge factual reliability and imaginative versatility. It thus provides theoretical depth and a unique method to evaluate hallucination in LLMs.