Stand-Up and Deliver is a MSc by Research Thesis from a student at the University of Kent looking at the applications of modern Large Language models in the domain of stand-up comedy routine generation.
This MSc by Research takes place between September 2022 and December 2024, with revisions from October 2025 to April 2026.
Author: Tommy Hills
Supervisor: Anna Jordanous
<aside> <img src="/icons/list-indent_blue.svg" alt="/icons/list-indent_blue.svg" width="40px" /> Table of Contents
</aside>
<aside> <img src="/icons/link_green.svg" alt="/icons/link_green.svg" width="40px" />
Link to Thesis on the Kent Academic Repository
</aside>
<aside> <img src="/icons/book-closed_purple.svg" alt="/icons/book-closed_purple.svg" width="40px" /> Abstract of the Thesis
This thesis investigates the self-improvement capabilities of Large Language Models (LLMs) through the domain of stand-up comedy. With its fundamental reliance on continuous feedback and iterative refinement, stand-up presents an ideal medium for exploring whether computational systems can demonstrate genuine creative evolution. Specifically, this research focuses on the generation of long-form anecdotes as part of a cohesive narrative routine, moving beyond the traditional computational focus on short-form wordplay.
To achieve this, the study introduces a novel prompt-engineering methodology designed to guide several commercial LLMs through a simulated “Engagement-Reflection" cognitive model. This pipeline forces the models to sequentially draft, self-critique, and iteratively improve comedy material. Throughout this process, the models demonstrate the capacity to identify structural flaws and independently apply revisions that align with established comedic mechanics, such as comedic timing, incongruity, and known techniques of the medium.
An analysis of the results reveals that clear, actionable improvements are both identified and executed by the models. By employing a triangulation evaluation strategy, the research highlights substantial quantitative and qualitative differences between the iterative drafts, showcasing measurable changes in pacing, semantic cohesion, and the underlying structure of the humour.
This thesis presents a promising framework for interacting with LLMs to reach an optimal creative result. The generated material exhibits clear “Progression and Development”, fulfilling a key component of computational creativity, and suggests significant applications for these models within the broader creative writing industry. However, the ultimate success of the generated routines as viable stand-up material remains unclear. Determining true performative validity requires further evaluation by domain experts or through live stage performance.
</aside>
<aside> <img src="/icons/code_blue.svg" alt="/icons/code_blue.svg" width="40px" /> Datasets used in this Thesis
</aside>
<aside> <img src="/icons/chart-area_red.svg" alt="/icons/chart-area_red.svg" width="40px" />
Analysis used in this thesis
Levenshtein Edit Distance per Generation
Levenshtein Edit Distance per LLM model
Semantic Distance across Generations
Joke-by-Joke Semantic Distances
Survey Results (Wave 1) with Semantic Distance Scoring
</aside>
$$ \textsf{University of Kent} $$