Generating Clue-Driven Investigative Game Narratives with Large Language Models
Creating interactive narratives where player investigation meaningfully drives story progression remains a central challenge in game design. We present a procedural narrative generation framework that leverages large language models (LLMs) to construct clue-driven investigative stories that unfold through character interactions and environmental exploration in a rich 3D game world. At the heart of the framework is a deductive solution model that defines what the player must ultimately infer through conversations with non-player characters (NPCs) and by examining in-world artifacts such as books and posters. This model guides the dynamic generation of characters, relationships, clues, and interactive dialogue to maintain narrative coherence. The framework automatically assembles the generated elements into a playable 3D episode in which players explore narrative elements that progressively reveal and incrementally narrow the space of valid interpretations, culminating in a final deductive conclusion. We evaluate the framework using automated tests of solvability and narrative consistency, as well as player experience studies assessing engagement, clarity, and satisfaction with the investigative gameplay. Results demonstrate that this approach can generate coherent, playable investigative narratives across a broad range of themes and subject matter areas (e.g., STEM, logistics, analytical thinking), enabling more scalable authoring of interactive story-driven games.