LAWRENCE LEROY TZE YAO CHIENG
Chan Chee Seng
Dr. Fan Lixin
This research
project delves into the phenomenon of 'hallucination' in Large Language Models
(LLMs), with a particular focus on OpenAI's ChatGPT. Hallucination is
characterized by the generation of outputs that lack factual grounding or
deviate from the provided prompt. The primary objectives of this study are to
gain a comprehensive understanding of ChatGPT, explore the occurrence of
hallucination in LLMs, and devise an effective solution to mitigate this issue.
The research encompasses an examination of various facets of LLMs, including
the latest GPT-4, LLM Augmenter, Reinforced Learning with Human Feedback
(RLHF), Black-Box Hallucination Detection, and a Probabilistic Model of Hallucination.
It also scrutinizes the token limits of these models. The study identifies a
significant gap in current research, namely the insufficient exploration of the
fundamental principles of hallucination and the computational demands of output
quality checks. The proposed methodology involves a detailed analysis of
hallucination types and the development of techniques to curb hallucination,
with a focus on English language content. The ultimate goal of the project is
to propose a solution that not only effectively curbs hallucination but also
enables LLMs to process extensive domain-specific knowledge, all while ensuring
computational and time efficiency.