How to Craft Fake Content Efficiently: Exploring Diversity in Large Language Model Outputs

Abstract

With the increasing importance of language models in research, investigating specific quality metrics is crucial. This study examines the diversity of outputs generated by them under varying conditions, focusing on temperature and the number of outputs per generation. Findings indicate that higher temperatures and increased number of outputs generally result in more unique and diverse outputs, while single outputs show higher convergence at low temperatures. Lexical and semantic similarities decrease with rising temperatures, increasing diversity but potentially reducing coherence. Text length is affected by temperature and output type, creating multiple outputs producing more concise texts and single outputs generating longer texts. Multiple output crafting is also more time-efficient than single one, making it suitable for applications needing diverse, quickly generated content. Insights are applicable to various fields, including creative content generation and the production of diverse, coherent, fake content, illustrating the complex dynamics of output diversity.