TLDR: This research introduces a novel framework for procedural game level generation guided by emotional arcs. By leveraging large language models, the system creates branching story graphs where each node is assigned an emotional trajectory (Rise or Fall), influencing gameplay elements like character attributes and difficulty. A prototype action RPG demonstrated that integrating emotional arcs significantly enhances player enjoyment and narrative coherence, and these arcs are perceivable by players and computational models. The work highlights the potential of emotionally structured procedural generation for advancing interactive storytelling in games, while also pointing to challenges in fully integrating narrative and gameplay mechanics.
In the evolving landscape of video games, narratives play a crucial role in shaping player experiences. While large language models (LLMs) have made significant strides in generating dynamic storylines and dialogue, a common challenge remains: ensuring these generated narratives possess global coherence and emotional depth. A new research paper, titled All Stories Are One Story: Emotional Arc Guided Procedural Game Level Generation, introduces an innovative framework that addresses this by integrating emotional arcs directly into the procedural generation of game narratives and levels.
The core idea behind this research is that all stories, across cultures and media, share universal narrative structures, often characterized by emotional arcs—patterns of rising and falling affective tone. Drawing on established narratological theories and large-scale empirical analyses, the researchers focus on two fundamental emotional patterns: Rise and Fall. These patterns serve as a structural backbone, guiding both story progression and gameplay dynamics.
How the Framework Works
The system begins by allowing a user to input a concise narrative prompt and select one of six predefined emotional arc types, such as ‘Rags to Riches’ (Rise) or ‘Tragedy’ (Fall). Using an advanced language model, GPT-4o-mini, the system then generates a branching story graph. Each node in this graph is assigned an emotional label—either Rise or Fall—indicating its intended narrative trajectory. The edges between nodes define the criteria for progression, like ‘talk to a friendly NPC’ or ‘defeat an enemy’.
A unique aspect of this framework is its human-in-the-loop customization. After the initial graph generation, designers can visualize and revise the emotional labels, storylines, and transition criteria through an interactive web interface. This ensures that the generated narrative aligns with creative intent. Once finalized, the system refines the textual content of each story node, injecting the emotional essence of the chosen arc, transforming structural labels into concrete narrative expressions.
Generating Game Worlds and Entities
Beyond just stories, the framework automatically converts each story node into a corresponding game level. This involves extracting and instantiating game-relevant entities such as characters (NPCs, enemies), items, and gameplay-relevant attributes like health and attack. A central design objective is to align gameplay difficulty with the emotional trajectory. For instance, ‘Rise’ nodes are designed to be more optimistic and approachable, while ‘Fall’ arcs introduce heightened tension and challenge, reflected in enemy statistics.
To maintain narrative consistency, the system considers the full history of prior nodes and generated entities. If a character acquires a new item in one level, that state change is reflected in subsequent levels, ensuring logical coherence and smooth gameplay transitions. For visual representation, the system integrates a diffusion model to generate pixel-art style sprites for game entities, which can be directly imported into game engines like Unity.
A Prototype Action RPG
To evaluate the practical application of their framework, the researchers implemented a prototype action role-playing game (ARPG). In this prototype, each node in the generated story graph corresponds to a self-contained dungeon room. Game entities are placed within these rooms based on the story description, and navigation to subsequent nodes is managed through doors, often gated by specific in-game interactions. This setup effectively translates the branching story graph into an interactive game experience.
Evaluation and Results
The framework was evaluated through a mixed-method approach, including a user study with 16 participants and automated sentiment analysis. The study aimed to answer two key questions: Does embedding emotional arcs enhance player experience? And are emotional arcs perceptible to players and computational models?
The results were compelling. Participants reported significantly higher enjoyment in games generated with emotional arcs compared to baseline versions without them. A large majority of participants (81.25%) could correctly identify the emotional arc condition, indicating that the arc structure was clearly perceivable. Qualitative feedback from interviews further supported these findings, with users describing arc-based stories as “more dramatic,” “better written,” and “more realistic.”
Automated sentiment analysis also validated the system’s ability to produce emotionally coherent story progressions. The valence trajectories (emotional tone) of the generated stories closely mirrored the canonical shapes of the six emotional arcs, especially for complex non-monotonic forms like ‘Rise-Fall-Rise’.
Also Read:
- The AI Evolution in Virtual Worlds: How Language Models are Redefining VR Gaming
- Advancing Emotion Understanding with Multimodal AI: A Deep Dive into Language Models
Challenges and Future Directions
While the findings affirm the efficacy of emotional arcs in enhancing story perception, the study also highlighted areas for improvement. Differences in perceived relevance and difficulty were not statistically significant, suggesting a need for tighter coupling between narrative structure and gameplay mechanics. Some participants noted that arc-generated narratives could be overly verbose, interrupting gameplay rhythm, and that gameplay interactions sometimes lacked depth.
Future work aims to address these limitations by extending the framework to other game genres, integrating non-verbal emotional signals (audio, visuals, pacing), and developing personalized narrative scaffolding based on player preferences. This research paves the way for more coherent, immersive, and emotionally resonant interactive storytelling experiences in games, demonstrating the significant potential of emotionally structured procedural generation.
