SNAP: A Plan-Driven Framework for Controllable Interactive Narrative Generation

Geonwoo Bang, DongMyung Kim, Hayoung OharXiv 2026 · DOI · arXiv

TL;DR

This paper introduces SNAP, a novel framework for controllable interactive narrative generation using Large Language Models (LLMs). SNAP addresses the challenge of LLMs drifting from narrative consistency by segmenting stories into "Cells" and employing explicit "Plans" to guide dialogue, ensuring plot progression and character consistency even with user-driven deviations.

What Problem Does This Paper Address?

Large Language Models (LLMs) have shown immense promise for creating engaging web-based interactive experiences like games and digital storytelling. However, a significant hurdle is their tendency to lose narrative consistency, especially in conversational settings where users can introduce unexpected inputs. LLMs can exhibit "spatiotemporal distortions," referencing future events or places not yet introduced, thus breaking immersion and deviating from the creator's intended narrative. Existing solutions like character-specific safeguards or memory augmentation have limitations, and prompt engineering or fine-tuning struggle to generalize across diverse narratives. This paper tackles the problem of maintaining narrative coherence and creator intent in interactive storytelling when faced with dynamic user interactions.

Methods

SNAP introduces two core concepts for controllable narrative generation:

  • Cell-Based Narrative Segmentation: The narrative is broken down into discrete, self-contained units called "Cells," akin to scenes. Each Cell has a defined scope, limiting the LLM's context to the current segment. This prevents "spatiotemporal leakage" where the LLM might access information from future parts of the story. Cells also provide clear progress markers and allow for creator intervention points between them.
  • Plan-Driven Dialogue Generation: For each Cell, a "Planning Model" generates candidate "Plans." A Plan is a structured sequence of sub-plans that outline the required events, actions, locations, and constraints for that segment. This Plan acts as a guide for the dialogue agent, ensuring it adheres to the narrative's progression and creator's intent without stifling conversational naturalness. A "Planning Filter" selects the optimal Plan based on coherence, connectivity, and personality consistency. The "Agent Model" then generates dialogue, balancing conversational naturalness with narrative progression according to the chosen Plan.

Datasets

The paper uses the Wikiplot dataset, which contains over 112,000 narrative plots, for evaluating SNAP. They randomly sample 13 plots and segment them into Cells of approximately 10 sentences each. Prompts for each phase of the evaluation are automatically generated using GPT-40.

Key Insights and Ideas

  • Cellular structure is crucial for preventing spatiotemporal distortions: By limiting the LLM's context to a single Cell, the system naturally avoids referencing future events or distant locations, a common failure mode in long-context LLMs.
  • Explicit Plans are powerful control levers: Even without Cell segmentation, Plan-driven generation significantly improves narrative control, demonstrating the effectiveness of structured guidance for LLMs.
  • Human perception of controllability differs from automatic metrics: While automatic evaluations might show slight trade-offs in certain metrics (like Linearity), human evaluations reveal that SNAP's structural mechanisms are highly effective in managing real user diversions and maintaining narrative coherence.
  • Controllability vs. Exploration Trade-off: The structured nature of SNAP, while enhancing control, can sometimes be perceived by users as restrictive, highlighting an inherent tension between narrative adherence and player freedom.

Key Results

  • Automatic Evaluation: SNAP consistently outperformed Vanilla GPT-40 and SNAP w/o Cell in metrics like Continuity and Information Appropriateness across 13 scenarios. For instance, SNAP achieved an average Continuity score of 4.510 compared to Vanilla's 3.819.
  • Human Evaluation: Participants rated SNAP significantly higher than Vanilla GPT-40 across all controllability dimensions (Continuity, Information Appropriateness, Non-Redundancy, Linearity) with substantial percentage improvements (e.g., Continuity +57.7%, Information Appropriateness +42.8%).
  • Ablation Study: Removing Cell segmentation (SNAP w/o Cell) led to performance degradation in Continuity and Non-Redundancy, underscoring the importance of this structural component.

Limitations and Future Work

  • Perceived restrictiveness: Some users found SNAP to be "sometimes felt restrictive, less freedom to explore." This suggests a need for more nuanced control mechanisms that balance narrative adherence with player agency.
  • Future Work: The authors suggest that SNAP provides a foundation for controllable narrative generation, paving the way for more reliable and creator-aligned interactive experiences. Future work could explore more sophisticated ways to manage the trade-off between structural control and exploratory freedom.

Summary updated

BibTeX
@misc{bang2026snap,
  author = {Geonwoo Bang and DongMyung Kim and Hayoung Oh},
  title = {{SNAP: A Plan-Driven Framework for Controllable Interactive Narrative Generation}},
  year = {2026},
  eprint = {2601.11529},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.11529},
}
Figure