A Survey on LLMs for Story Generation

Maria Teleki, Vedangi Bengali, Xiangjue Dong, Sai Tejas Janjur, Haoran Liu, Tian Liu, Cong Wang, Ting Liu, Yin Zhang, Frank Shipman, James Caverlee

Abstract: Maria Teleki, Vedangi Bengali, Xiangjue Dong, Sai Tejas Janjur, Haoran Liu, Tian Liu, Cong Wang, Ting Liu, Yin Zhang, Frank Shipman, James Caverlee. Findings of the Association for Computational Linguistics: EMNLP 2025.

2025.

1. Introduction

Large Language Models (LLMs) have revolutionized text generation, yet creative storytelling remains a uniquely difficult "final boss" for artificial intelligence. Narrative generation requires more than linguistic fluency; it demands long-term logical consistency, emotional resonance, and a deep understanding of subtext—areas where LLMs frequently struggle with hallucinations or narrative drift. This survey provides a systematic overview of the "LLM era" of story generation (post-2022), organizing recent research into a coherent taxonomy to guide future development in the field.

2. A Taxonomy of Narrative Generation

The authors categorize recent advancements based on the primary driver of the creative process. This taxonomy distinguishes between systems where the AI acts as an autonomous creator and those where it serves as a collaborative tool for human authors.

Comprehensive taxonomy of LLM-based story generation systems, including single-agent, multi-agent, adaptive, and static approaches.

2.1 Independent Generation

In independent generation, the LLM is the primary author. This branch is further divided into:

  • Single-Agent Generation: Employs techniques such as constrained generation, direct prompting, or hierarchical planning. For instance, the DOME framework uses outline-based generation to maintain structure over long sequences, significantly outperforming "flat" prompting.
  • Multi-Agent Collaboration: Systems like COLLABSTORY and CRITICS utilize specialized LLM agents (e.g., planners, writers, and critics) to mitigate coherence issues. By assigning roles like a "Temporal Conflict Analyzer," these pipelines can catch plot contradictions before they reach the reader.

2.2 Author Assistance

In this paradigm, the human remains the lead author, utilizing the LLM as a creative partner. This includes Adaptive Stories (e.g., STORYVERSE), which evolve based on user interaction while adhering to high-level narrative frameworks, and Static Stories (e.g., CHARACTERMEET), which assist in specific creative tasks like character brainstorming or dream recording.

3. Authorship Paradigms

The distinction between independent and assisted generation is central to understanding the current research landscape. While independent systems focus on architectural improvements for coherence, assistance systems prioritize interactivity and the user's creative agency.

Conceptual comparison between Independent Generation (AI-led) and Author Assistance (Human-led) workflows.
  • From Strings to "Abstract Acts": Modern interactive systems are moving away from generating raw text strings. Instead, they generate "abstract acts" (e.g., Character X falls in love) that represent narrative beats, which are then translated into specific game or story actions to balance authorial intent with player freedom.
  • Educational Scaffolding: LLMs are being leveraged for pedagogical storytelling, such as MATHEMYTHS, which helps children learn math vocabulary by weaving it into collaborative narratives.
  • The Evaluation Crisis: The field currently faces a significant challenge in measurement. Traditional automated metrics like BLEU and ROUGE fail to capture narrative quality or creativity, while small-scale user studies (often N < 20) are difficult to generalize.

5. Challenges and Future Directions

Despite progress, several hurdles remain. Even with advanced memory modules, LLMs struggle to maintain global coherence over full-length novels. Furthermore, multimodal consistency remains a pain point; ensuring a character's visual portrayal remains stable across generated text and imagery is a major area for future work. The authors suggest exploring inference-time constraints (such as constrained beam search) to enforce narrative structure without the need for extensive retraining, while also emphasizing the need for transparency regarding the ethical implications of AI in creative professions.

Summary updated

BibTeX
@inproceedings{teleki2026survey,
  author = {Maria Teleki and Vedangi Bengali and Xiangjue Dong and Sai Tejas Janjur and Haoran Liu and Tian Liu and Cong Wang and Ting Liu and Yin Zhang and Frank Shipman and James Caverlee},
  title = {{A Survey on LLMs for Story Generation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2026},
  url = {https://openreview.net/forum?id=HvvL6BwbbQ},
}
Figure