Player-Driven Emergence in LLM-Driven Game Narrative

TL;DR

This work investigates how players can catalyze emergent narrative elements by interacting with GPT-4 powered NPCs in a text-adventure (Dejaboom!). Using GPT-4 to convert gameplay logs into narrative graphs and comparing them with designer-authored graphs, the authors identify emergent nodes—novel narrative elements not present in the original design—arising from non-deterministic LLM output. Across 28 participants, 53 emergent nodes (43 unique) emerged, spanning information extraction, new entities, hidden information, and creative defusal strategies, with emergence correlating to certain player motivation profiles. The findings suggest a collaborative design paradigm where player creativity and LLM-driven dialogue expand the space of possible narratives, albeit with latency and consistency challenges to address in future work, and point toward leveraging emergent feedback for richer open-world game development.

Abstract

We explore how interaction with large language models (LLMs) can give rise to emergent behaviors, empowering players to participate in the evolution of game narratives. Our testbed is a text-adventure game in which players attempt to solve a mystery under a fixed narrative premise, but can freely interact with non-player characters generated by GPT-4, a large language model. We recruit 28 gamers to play the game and use GPT-4 to automatically convert the game logs into a node-graph representing the narrative in the player's gameplay. We find that through their interactions with the non-deterministic behavior of the LLM, players are able to discover interesting new emergent nodes that were not a part of the original narrative but have potential for being fun and engaging. Players that created the most emergent nodes tended to be those that often enjoy games that facilitate discovery, exploration and experimentation.

Figures (3)

• Figure 1: Game Design: A map of the village where the game takes place, showing the locations, objects and NPCs. The player begins the game from home and their goal is to defuse the bomb before it explodes again.
• Figure 2: Our implementation of the game. We use TextWorld as our fixed game agent. The user's command is first classified by GPT-4 as [action] or [words]. If an [action] (e.g. "chase the birds"), then it is passed on to the fixed game agent that either executes the action or returns a failure message (e.g. "This verb is not recognized") which is further processed by GPT-4 to generate appropriate game feedback (e.g. "You tried to chase the birds, but nothing happened"). If [words] for an NPC (e.g. "can I see your menu"), then the command is passed to GPT-4 , which uses the game history and the NPC's information to generate an appropriate NPC response (e.g. "Chef Maria: Of course! Our menu today features a delicious selection of Italian-American fusion dishes.")
• Figure 3: The blue nodes make up the initial narrative graph intended by the designers. A single path from the start node to the end node indicates one possible way to stop the explosion. The green nodes are a few examples of the emergent nodes created by players implicitly through their gameplay.