Diversity in Language Model Generation
Studying how large language models select from multiple valid outputs and why diversity collapses in generation.
This project studies diversity in language model generation, especially when multiple outputs are valid but models systematically prefer a small subset of them.
We investigate how factors such as prompt framing, output ordering, temperature, and validity constraints affect model choices. The goal is to better understand when diversity matters for utility and how generation systems can be designed to avoid unnecessary collapse to a narrow set of outputs.
Keywords: Large language models, diversity, generation, model behavior, human-centered AI