UC Berkeley

Large language models (LLMs) have become remarkably proficient at generating coherent, high-quality responses. However, they often struggle to produce diverse outputs, especially when multiple, equally plausible answers exist. This limitation becomes critical in areas like adversarial testing, search, and synthetic data generation, where generating distinct yet valid responses is essential. This thesis opens with a discussion of the timely need for addressing the diversity of language models. I’ll then introduce SimpleStrat, a lightweight method that leverages the model itself to automatically stratify the solution space. By performing stratified sampling, we can improve diversity and coverage without sacrificing quality. I’ll also describe how we measure resampling diversity using CoverageQA, and discuss how these ideas extend to creative and open-ended tasks. To extend these ideas into the weight space, I’ll discuss Stylus, which navigates the quality-diversity tradeoff by leveraging adapter merging to produce higher-quality images while accessing meaningful diversity by randomizing subsets of adapters for merging. The thesis will conclude with a discussion on going beyond diversity to creativity and current challenges in applying LLMs to discover novel and creative work. Anyone can critique a masterpiece, but immortal is the artist who creates one.