The output of large language models (LLMs), even across model families, is becoming increasingly homogeneous…. We generate 20,000 stories with four current models from OpenAI, Anthropic, Google, and the Allen Institute for AI (AI2), finding 88.3% of generated stories contain one of 11 core words (including character names, story locations, and professions). Most notably, over half feature a lighthouse. Why does this story pattern become favored? These 11 words are not common in published English literature, which suggests that post-training data is responsible. But examining OLMo 3’s post-training sets reveals only 3,053 out of a total 78,958 stories contain one of our 11 words. We find that the dominance of “Elias in the Lighthouse” stories cannot be explained by prevalence in pre- or post-training data.
from “Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories” by Sil Hamilton and David Mimno
The 11 words are “Lighthouse,” “Mara,” “Elias,” “Elara,” “Keeper,” “Baker,” “Mayor,” “Clockmaker,” “Fisherman,” “Librarian,” and “Conductor.”
