StorySpark: Evolutionary Search for Story Premise Generation
A new framework named StorySpark has been unveiled by researchers for crafting story premises through a module-wise evolutionary search method. This framework, outlined in a paper on arXiv (2608.12336), focuses on the lesser-explored domain of premise-level ideation in story generation using LLMs. StorySpark utilizes interpretable narrative components like background, persona, event, ending, and twist, treating each as a localized search area based on the existing partial premise. It produces alternatives, assesses them contextually, and enhances them via feedback-driven mutation and recombination, employing Pareto-guided selection to maintain complementary strengths. Additionally, it reallocates frontier capacity to optimize branch coverage with promising paths. The paper features multi-view evaluations, although specific findings are not included in the abstract, emphasizing a shift towards more innovative and exploratory facets of AI storytelling.
Key facts
- StorySpark is a module-wise evolutionary search framework for story premise generation.
- It operates over narrative modules: background, persona, event, ending, and twist.
- Each module is treated as a local search space conditioned on the partial premise.
- The framework uses feedback-driven mutation and recombination for refinement.
- Pareto-guided selection preserves complementary strengths.
- It reallocates frontier capacity to balance branch coverage and promising directions.
- The paper is available on arXiv with ID 2608.12336.
- The work emphasizes premise-level ideation, an underexplored area in LLM story generation.
Entities
Institutions
- arXiv