Aligning LLMs for Poetic Expression
In this white paper explore the development of FredTheHeretic (FTH), an LLM-based system for generating poetry in the style of Fred Turner. We examine the challenges of evaluating AI-generated poetry and propose alternative assessment frameworks beyond traditional NLP metrics. While implementing FTH as a Streamlit application, we discovered that conventional evaluation metrics fail to capture the essence of poetic quality. We review existing approaches to poetry evaluation from literature and propose a novel graph-based analysis method that maps semantic richness. This paper serves as a foundation for future work on memory frameworks that enable persistent learning across sessions and more sophisticated evaluation techniques for AI poetry
Lead Author: Mihir Dattatraya Hirave