narrative design

This work addresses the fundamental tension in interactive storytelling between authorial intent and player agency. It proposes Dramamancer, a novel design paradigm for large language model (LLM)-driven interactive narratives that reconciles these competing demands by parsing author-defined story schemata and integrating dynamic response mechanisms to player actions. Leveraging the generative capabilities of LLMs, the system produces narrative content in real time that adheres to underlying story logic while supporting high degrees of interactivity. Empirical evaluation demonstrates that this approach effectively balances narrative coherence with creative control and player autonomy. The study thus offers a reproducible design framework and evaluation pathway for LLM-powered interactive storytelling systems.

This study investigates how to support players’ reflection on socialization processes, moral responsibility, and educational roles within AI-mediated environments. To this end, we designed a large language model–based role-playing game that guides players through a “four seasons” narrative structure, accompanying a virtual prince across staged moral dilemmas. The system innovatively incorporates delayed feedback and phased growth prompts to balance open-ended expression with sustained engagement. By deliberately omitting immediate scoring and emphasizing reflective scaffolding, this work presents the first implementation of socialization theory into an interactive AI game system. A user study (N=12) demonstrates that players effectively negotiate their sense of responsibility and role identity, while also revealing a tension between openness in expression and the entry load required for continued participation.

This study addresses the tension between factual accuracy and engaging interaction when deploying large language models (LLMs) as role-playing chatbots in museum settings, where enhancing historical fidelity often compromises conversational liveliness. The authors propose a “critical play” design paradigm that reframes the inherent unreliability of LLMs as a pedagogical opportunity, positioning the model as a fictional character embodying historical narratives, multiple perspectives, and ironic commentary. By integrating narrative design with interactive strategies, the system fosters audience engagement while prompting critical reflection on the complexities of historical representation. This approach offers a novel and responsible pathway for applying LLMs in cultural heritage contexts, transforming potential limitations into meaningful educational experiences.

This paper addresses the insufficient immersion and sense of agency in LLM-driven interactive drama by proposing a novel framework to enhance player narrative agency. Methodologically: (1) it introduces scriptwriting-guided generation—a first-of-its-kind approach that explicitly constrains narrative structure to improve plot coherence; and (2) it designs an LLM agent reflection mechanism grounded in a narrative graph, enabling precise player intent recognition and responsive adaptation. The framework integrates character agent modeling, real-time dialogue interaction, and human-in-the-loop evaluation for iterative optimization. Extensive multi-round experiments demonstrate significant improvements over baseline systems: +32.7% in immersion, +41.2% in agency, +28.5% in character consistency, and markedly increased player influence over plot progression. This work establishes a reproducible technical pathway toward trustworthy, controllable, and highly engaging AI-driven narrative experiences.

Large language models (LLMs) exhibit limited high-level strategic planning capabilities in complex multi-agent games, while conventional reinforcement learning (RL) approaches suffer from heavy reliance on extensive training data. Method: This paper proposes a novel two-level tree search–driven LLM self-play learning paradigm that synergistically integrates Monte Carlo Tree Search (MCTS) with LLM-based reflective reasoning: strategic-level state evaluation and policy planning, and execution-level action generation and dialogue synthesis. End-to-end strategic skill acquisition is achieved via self-play and reinforcement feedback. Contribution/Results: To our knowledge, this is the first framework to bridge the gap between LLMs and symbolic decision-making. It significantly improves win rates and strategic robustness on benchmark games—including GOPS and *The Resistance: Avalon*—outperforming both standard RL baselines and state-of-the-art LLM skill-learning methods.

This study investigates how large language models (LLMs) are reshaping game design paradigms and influencing gameplay, playability, and player experience. Through two LLM-integrated game projects, the research employs a collaborative autoethnographic approach, combining development practice, reflective narratives, and qualitative content analysis to systematically examine the role of LLMs within game architectures. The work provides the first empirical evidence that LLMs significantly enhance content diversity and personalization, yet simultaneously introduce novel challenges concerning factual accuracy, difficulty balancing, and narrative coherence. Building on these findings, the study proposes new architectural principles and quality evaluation dimensions specifically tailored for generative AI–driven games.

This work addresses the challenge educators face in effectively mapping instructional objectives to game mechanics, a process often hindered by high design barriers and opaque AI recommendations in existing tools. To overcome these limitations, the authors propose a web-based co-design tool grounded in controlled natural language, which enables teachers to collaboratively construct structured mappings between pedagogical intent and gameplay mechanics through an explicit, editable linguistic interface. The system employs a four-component interactive architecture that preserves human agency via a human-in-the-loop design while enhancing alignment between educational goals and game dynamics. Empirical evaluation demonstrates that the tool significantly lowers the entry barrier for non-expert developers, improves their comprehension of and control over AI-assisted design processes, and effectively facilitates the coherent integration of learning outcomes with game mechanics.

This study investigates how prompt constraint strength (high vs. low) affects non-scripted NPC dialogue experiences in LLM-driven games. To address the trade-off between coherence and improvisational expressiveness, we propose the “Symbolic Scaffolding Game” framework, which employs boundary-blurring symbolic design to generate role-sensitive prompts—challenging the prevailing assumption that stronger constraints inherently yield superior outcomes. Integrating GPT-4o, JSON-structured prompting, and RAG, we conduct empirical evaluation within a voice-based detective game, employing both LLM-based adjudication for synthetic assessment and within-subject usability testing. Results reveal role-dependent constraint effects: high constraint improves behavioral stability for task-giving NPCs (e.g., investigators) but significantly degrades conversational naturalness for suspect NPCs. This validates the efficacy and necessity of role-adaptive scaffolding mechanisms in generative NPC dialogue systems.

Addressing the challenge of jointly optimizing global narrative structure and local textual details—and the resulting consistency degradation from multi-granularity revisions—in long-form script editing, this paper proposes a hierarchical multi-agent collaborative framework. The method employs a top-down task flow with a coarse-to-fine iterative mechanism, comprising three sequential stages: global narrative review, scene-level review, and hierarchical coordinated revision, thereby enabling sustained high-level intent guidance over local modifications. The framework supports plug-and-play integration of existing generative models, preserving contextual coherence and narrative logic. Experimental results demonstrate statistically significant improvements over baseline models across key dimensions—including overall script coherence, plot plausibility, and descriptive richness—indicating enhanced artistic quality and structural integrity in generated scripts.

Current LLM-driven interactive drama systems lack a unified, reproducible development platform, hindering research progress and system scalability. To address this, we introduce DramaLLM—the first open-source framework specifically designed for interactive drama. It features a modular, configurable multi-agent architecture integrating a hierarchical retrieval-augmented memory mechanism and an LLM-powered dialogue control pipeline, significantly improving character behavioral consistency and long-horizon narrative coherence. The framework supports immersive user role-playing and dynamic plot evolution, while ensuring high extensibility and experimental usability. Empirical evaluation across diverse dramatic scenarios demonstrates robust system stability alongside strong narrative expressiveness. DramaLLM thus establishes a standardized, accessible infrastructure for advancing interactive narrative research.