Creating Suspenseful Stories: Iterative Planning with Large Language Models

TL;DR

The paper tackles the challenge of generating suspenseful stories with large language models by introducing a theory-grounded, zero-shot iterative prompting method called Iterative-Prompting-based Planning. It leverages cognitive psychology and narratology to structure story development into Background Setup, Outline Planning, and Detail Elaboration, without relying on supervised story corpora. Through extensive human evaluations and ablation studies, the approach demonstrates increased suspense, novelty, and coherence compared to baselines, and analyzes how clues, information revelation timing, and reader empathy influence suspense perception. The method shows transferability across LLMs (e.g., ChatGPT and Llama 2) and contributes to a deeper understanding of the mechanisms that drive suspense in generated narratives.

Abstract

Automated story generation has been one of the long-standing challenges in NLP. Among all dimensions of stories, suspense is very common in human-written stories but relatively under-explored in AI-generated stories. While recent advances in large language models (LLMs) have greatly promoted language generation in general, state-of-the-art LLMs are still unreliable when it comes to suspenseful story generation. We propose a novel iterative-prompting-based planning method that is grounded in two theoretical foundations of story suspense from cognitive psychology and narratology. This theory-grounded method works in a fully zero-shot manner and does not rely on any supervised story corpora. To the best of our knowledge, this paper is the first attempt at suspenseful story generation with LLMs. Extensive human evaluations of the generated suspenseful stories demonstrate the effectiveness of our method.

Figures (2)

• Figure 1: Our suspenseful story generation technique is predicated on a theory of suspense in which the quantity and quality (likelihood) of ways in which a protagonist can avoid an undesirable outcome are decreased. Our method iteratively produces possible actions for the protagonist to take and adversarially creates the conditions under which the protagonist will fail. The dialogue boxes indicate the story snippets generated by our method. The thought bubbles of the reader represent the mental process. The thought bubbles of the system depict the planning process in our method.
• Figure 2: Human evaluation results for the investigation on reader's empathy as visualized in a heatmap. The numbers in the heatmap indicate the counts of cases in which human participants feel the corresponding levels of empathy and suspense after reading a suspenseful story.