NarrativeWorlds: A Framework for Authorial Control and Game-State Consistency in LLM-Driven Interactive Digital Narratives

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


MPhil Thesis Defence


Title: "NarrativeWorlds: A Framework for Authorial Control and
Game-State Consistency in LLM-Driven Interactive Digital Narratives"

By

Mr. Serkan KUMYOL


Abstract:

Large language models (LLMs) are increasingly used in interactive digital 
narratives (IDNs) for both authoring and character interaction. Their ability 
to interpret and generate open-ended language expands the range of possible 
author requests, player actions, and NPC responses, while also creating 
challenges for narrative control. Model-produced dialogue and interpretations 
may imply changes to characters, relationships, secrets, or events whose 
authority within the fiction remains unclear. This thesis therefore 
investigates how such information can be represented, checked, and given 
bounded authority within an interactive narrative system.

We develop NarrativeWorlds, a framework that examines this problem across 
authoring, social interaction, structured character state, and runtime world 
change. Narrative Anvil translates natural-language author requests into 
typed, structurally validated operations over a persistent narrative graph, 
with traceability and turn-level reversal support. NarrativeHive studies 
staged social reasoning through Perception, Stance, Opinion, and Response, 
together with a longer-horizon reputation mechanism. NarrativeSignals 
represents an NPC's social situation as a 29-field structured state predicted 
before response generation and used for inspection and deterministic routing. 
NarrativeTown compiles the authored world into a playable simulation and 
represents dialogue-implied changes as candidate updates governed before they 
enter authoritative state. Narrative Anvil and NarrativeTown form the 
implemented authoring-to-runtime path, while NarrativeHive and 
NarrativeSignals are evaluated independently.

Across automated tests, model comparisons, a player study, annotation 
analyses, and simulated game sessions, the results show both the value and 
cost of making intermediate decisions explicit. NarrativeHive improved 
player-rated Social Presence at greater latency, while NarrativeSignals 
revealed construct-validity challenges through limited agreement with human 
judgements. Narrative Anvil demonstrated structural filtering, traceable 
world editing, and turn-level reversal on the tested paths, while 
NarrativeTown demonstrated governed handling of dialogue-implied consequences 
on the exercised runtime paths. Taken together, the thesis presents 
NarrativeWorlds as an architectural approach in which LLMs support flexible 
authoring and interaction while consequential narrative decisions remain 
connected to explicit representations and bounded decision authority.


Date:                   Wednesday, 26 August 2026

Time:                   2:00pm - 4:00pm

Venue:                  Room 2132C
                        Lift 22

Chairman:               Prof. Pedro SANDER

Committee Members:      Dr. Tristan BRAUD (Supervisor)
                        Prof. Andrew HORNER