Reducing LLM hallucinations: exploring the efficacy of temperature adjustment through empirical examination and analysis

• Main Author: Tan, Max Zheyuan
• Other Authors: Jun Zhao
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Hallucination; Foundation model
• Online Access: https://hdl.handle.net/10356/174789
• Institution: Nanyang Technological University

Foundation models are gaining widespread prominence both in industrial applications and among individual users. In the tech landscape today, prompt engineers play a crucial role by crafting industry-standard prompts, empowering companies to enhance productivity, engage with customers, and automate various tasks. At the same time, individuals leverage publicly available foundation models such as OpenAI's ChatGPT for everyday activities like summarising text or composing emails. However, a significant challenge to the reliability and usability of these models arises in the form of hallucinations, wherein the model generates content that is fundamentally incorrect. To address this issue, a prevalent approach involves minimising hallucinations by adjusting the sampling randomness of the foundation model during execution, achieved through the manipulation of the temperature hyperparameter. This paper delves into the validity of this method and explores its associated caveats. The findings are subsequently distilled into comprehensible insights applicable to a diverse audience. Moreover, the paper outlines potential avenues for future research, offering a foundation for further exploration in this domain.