Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories

2026-05-25

AI Summary

This study addresses the severe content homogenization in stories generated by large language models, characterized by repetitive use of specific words and plot elements. By systematically sampling 20,000 stories produced by four leading models across five prompts, the authors combine word frequency analysis, data provenance tracing, and comparisons between pretraining and post-training token distributions to demonstrate— for the first time—that small-scale preference datasets, when amplified through strong alignment algorithms, can induce significant output uniformity. The analysis reveals that 11 keywords (e.g., “Elias,” “lighthouse”) appear in 88.3% of generated stories, with frequencies vastly exceeding those in both real-world literary corpora and pretraining data, thereby exposing a hidden cost of current alignment mechanisms: the substantial erosion of textual diversity.

Abstract

LLM-generated stories are a popular use case, but they show very low variability. We sample 20,000 total stories from four current models using five prompts. We find that 11 words occur in 88.3% of generated stories, with little difference between models. These words include names (Elias, Mara, Elara), settings (lighthouses), and professions (clockmaker, librarian). These tokens do not often occur in published literature nor pre-training data, but they are found in preference data that is likely to have been used by all current models. Surprisingly, these "lighthouse" stories are infrequent when compared with the average post-training story, much of which contains references to copyrighted characters or adult content. This result demonstrates the potentially disproportionate impact of small datasets combined with powerful alignment algorithms.

Problem

• low diversity
• LLM-generated stories
• repetitive tokens
• preference data
• alignment algorithms

Innovation

• low diversity
• LLM-generated stories
• preference data
• alignment algorithms
• token repetition