Narrative Adherence in LLM-driven Games

Jørgensen, Nicolai Hejlesen ; Tharmabalan, Sarmilan

4. term

Software, Master

2025

2025-06-19

14

Dette speciale præsenterer ChatRPG v3, en AI der fungerer som spilleder i enspiller-rollespil. Vi introducerer SENNA, et system med fem samarbejdende agenter, der hjælper AI’en med at følge strukturerede, forhåndsskrevne eventyrmoduler og samtidig bevare meningsfulde valg for spilleren. Systemet bruger en fortællingsgraf—et kort over scener og forbindelser—til at holde styr på, hvor spilleren befinder sig i historien. Når spilleren bevæger sig væk fra den planlagte kurs, anvender SENNA omdirigeringsstrategier, der giver mening i spiluniverset, så spilleren ledes tilbage uden at bryde fiktionen. Eksempler omfatter små skub fra ikke-spiller-figurer (NPC’er) og troværdige konsekvenser i verdenen. Vi evaluerede SENNA med en brugerundersøgelse, der kombinerede live-spil, sammenligninger af alternative scener og spillerfeedback. Resultaterne viser, at struktureret, verdenslogik-baseret omdirigering forbedrer fortællingens sammenhæng uden at mindske indlevelse eller følelse af autonomi. Disse fund udvider designrummet for AI-drevet interaktiv fiktion og giver praktiske indsigter i, hvordan man balancerer skrevne historier med uforudsigeligt spillerinput.

This thesis presents ChatRPG v3, an AI that acts as a game master for single-player role-playing games. We introduce SENNA, a system of five cooperating agents that helps the AI follow structured, prewritten adventure modules while still allowing meaningful player choices. The system keeps a narrative graph—a map of scenes and connections—to track where the player is in the story. When players drift away from the planned arc, SENNA applies redirection strategies that make sense within the game world, guiding them back without breaking the fiction. Examples include nudges from non-player characters (NPCs) and believable in-world consequences. We evaluated SENNA through a user study combining live gameplay, comparisons of alternative scenes, and player feedback. Results show that structured, world-logic-based redirection improves narrative coherence without reducing immersion or a sense of autonomy. These findings expand the design space for AI-driven interactive fiction and offer practical insights on balancing authored stories with unpredictable player input.

[This summary has been rewritten with the help of AI based on the project's original abstract]

Agentic AI ; AI ; AI Game Master ; Artificial Intelligence ; C# ; ChatRPG ; ChatGPT ; Choose Your Own Adventure ; CYAO ; DM ; GM ; Dungeons and Dragon ; Dungeon Master ; Game Master ; Games ; Generative AI ; Narrative Adherence ; Narrative Redirection ; GPT-4o ; Interactive Fiction ; LangChain ; Large Language Models ; LLM ; Multi-Agent System ; Narrative ; OpenAI ; Player Experience Inventory ; PXI ; Artificial Social Agent Questionnaire ; ASAQ ; Prompt ; Prompting ; Prompt Engineering ; Role-Playing Games ; RPG ; Storytelling ; TTRPG ; User Engagement ; Narrative Graph ; ReAct ; SENNA ; Redirection Strategies ; Redirection ; Pre-written Adventure

Download
View record in AAU Student Projects