Where Models Converge and Humans Diverge: A Coverage Framework for Distributional Pluralism in Open-Ended Generation

The paper, arXiv:2608.05576, examines how LLMs reliably reproduce core elements of a universe (e.g., Hogwarts locations and characters) but fail to match the stylistic irregularity and relationship-diverse plotlines found in human-written fanfiction. This convergence-divergence pattern is observed across domains, suggesting a general limitation in current models. The proposed coverage framework offers a structured way to measure distributional pluralism, quantifying how much of the human output distribution is covered by model outputs. This can inform model evaluation and training objectives aimed at increasing output diversity. The research highlights the need for metrics beyond accuracy or coherence to capture creative variety.