Researchers have developed a metric called the Word Coverage Score (WCS) to measure how well large language models (LLMs) can generate diverse and unique text. The WCS assesses how sampling parameters in LLMs affect the likelihood of generating low-frequency words, which are often essential for conveying nuanced human expression. By applying the WCS to open-weight models on human-authored corpus fragments, the authors found that industry-standard sampling defaults can act as censorship mechanisms, suppressing the diversity of human language and producing repetitive text. --- Why it matters: Understanding how LLMs suppress linguistic diversity is crucial for optimizing their ability to generate coherent and rich text. The Word Coverage Score provides a diagnostic tool for researchers to identify areas where sampling parameters are inadvertently limiting the models' capacity for generating unique words, enabling them to improve the trade-off between text coherence and lexical richness. Source: https://arxiv.org/abs/2605.27268