Reproduce “How LLMs Build Fictional Worlds: Setting and Narrative Space in AI-Generated Creative Storytelling” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Katrin Rohrbacher, Björn Nieth, Emmanuelle Salin, Bjoern Eskofier, Michaela Mahlberg
Rohrbacher et al. investigate how large language models construct fictional worlds by analyzing narrative setting, revealing LLMs systematically overproduce 'perceived space' while often underrepresenting 'action space' compared to human authors. For instance, GPT 4.1 generated perceived space at roughly double the frequency of human texts in English story openings.
Key insights
- LLMs exhibit a distinct 'generative AI style' in worldbuilding, characterized by a systematic overemphasis on atmospheric 'perceived space' and often an underrepresentation of embodied 'action space'.
- LLM narrative patterns are shaped by both model characteristics and linguistic context, with open-source models aligning more closely with GPT 4.1 in English than in German.
- Iterative generation strategies and potential post-training alignment methods contribute to observed spatial stylistic biases, such as elevated perceived space at chapter boundaries.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.