Developing a Sense of Humour in Large Language Models

Artificial Intelligence, particularly Large Language Models (LLMs) such as GPT, Gemini, and Claude, demonstrate high levels of autonomy, sometimes even surpassing human capabilities across various domains. However, in the domain of comedy and humour, these models still show notable limitations: they often struggle to generate genuinely funny content, understand the subtleties of humour and are not effective at supporting the creative process involved in generating comedy.

This dissertation focuses on exploring the roots of these limitations, whether they are based on the lack of targeted training data, inherent architectural constraints, or other factors. With a particular emphasis on Portuguese humour, this research aims to investigate how LLMs can be improved to better understand and generate comedy content.

The core objectives of the thesis include

- Review related work, such as “A Robot Walks into a Bar: Can Language Models Serve as Creativity Support Tools for Comedy? An Evaluation of LLMs’ Humour Alignment with Comedians” ( https://arxiv.org/pdf/2405.20956 ) and define a roadmap for the thesis.

- Curate a dataset composed of comedy sketches and material from well-known Portuguese comedic groups (“Gato Fedorento”, “Herman José”, “Porta dos Fundos”…).

- Use this dataset in a Retrieval-Augmented Generation (RAG) pipeline to enhance LLM performance in comedy brainstorming and joke creation.

- Fine-tune open-source language models such as LLaMA, Mistral, or Qwen with the dataset to assess whether domain-specific tuning improves their humour capabilities.

- Analyze model activations and embeddings.

- Evaluate model outputs through structured human feedback to measure improvements in comedy quality and relevance.