Player-Driven Emergence in LLM-Driven Game Narrative

🌈 Abstract

The paper explores how interaction with large language models (LLMs) can give rise to emergent behaviors, empowering players to participate in the evolution of game narratives. The authors use a text-adventure game called "Dejaboom!" as a testbed, where players interact with non-player characters (NPCs) generated by GPT-4. The study analyzes the narrative paths developed by players and compares them to the original narrative intended by the game designers. The authors 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 part of the original narrative but have potential for being fun and engaging. The paper also explores the relationship between player motivation profiles and the creation of emergent nodes.

🙋 Q&A

[01] Game Design

1. What is the premise of the game "Dejaboom!"? The premise of the game "Dejaboom!" is that the player wakes up in their home with a sense of déjà vu, remembering an explosion that occurred in their village the previous day. The player is now reliving the same day with the goal of stopping the explosion.

2. What are the key locations and objects in the game? The game world consists of 6 main locations: a home, a park, a restaurant, a blacksmith's shop, a library, and a town hall. There are also two secret rooms - a secret lab in the library and a secret storage room in the blacksmith's shop where the bomb is hidden. The player can interact with actionable objects like ingredients for a bomb disposal kit and objects that provide clues.

3. What are the different types of NPCs in the game? The game features five NPCs - two well-known literary characters (Mad Hatter and Merlin) and three custom characters (Mrs. Thompson, Chef Maria, and Moriarty). The NPCs are strategically placed to provide the player with information and clues related to the bomb.

4. What are the main milestones the player needs to achieve to solve the game? The two main milestones are: 1) Locate the bomb, which is hidden in the blacksmith's storage room, and 2) Acquire the bomb disposal kit, which can be obtained by either collecting the required items or convincing Merlin to give it to the player.

[02] User Study

1. How many participants were recruited for the user study and what were their demographics? The authors recruited 28 participants (24 male, 3 female, 1 undisclosed) located in the United States to play the game.

2. How were the participants categorized based on their game motivation profiles? The participants were categorized into 6 main player motivation profiles: Action, Social, Mastery, Achievement, Immersion, and Creativity.

3. What were some of the key findings from the user study? Key findings include:

• 20% of players appreciated the flexible command choices and error-correction offered by the game
• 20% enjoyed the realistic and fun conversations with the NPCs
• 10% of players (and 100% of those motivated by creativity) appreciated the freedom to explore different parts of the game
• 54% of players found the fluidity of interactions with the NPCs to be the most enjoyable aspect
• 56% wanted to play again to find the bomb or solve the game

4. What were some of the areas for improvement identified by the players? Areas for improvement include:

• 40% were frustrated by latency issues
• 10% observed inconsistencies in NPC personas
• 10% saw repetitive responses from the NPCs
• 14.28% would have liked the NPCs to have a more mysterious character

[03] Emergence in Game Narrative

1. How did the authors use GPT-4 to generate narrative graphs for the players and the original game design? The authors used GPT-4 to convert the game logs of players and the game designers into narrative graphs, where each node represents a player's strategy at a certain point in the game, and the edges represent the temporal progression. The emergent nodes are those that appear in the player's narrative graph but not in the original narrative graph intended by the designers.

2. What were the main categories of emergent nodes identified in the player narratives? The main categories of emergent nodes include:

• Creative ways of extracting information from NPCs
• Suggestions for the addition of new entities (objects, locations, NPCs)
• Creative ways of finding hidden information
• New ways of defusing the bomb

3. How did the player motivation profiles relate to the creation of emergent nodes? The authors found that players with "mastery" and "creativity" motivation profiles tended to create the most emergent nodes, with the "creativity" profile having the highest nodes per player ratio.

4. What are the implications of the emergent nodes for game design? The authors suggest that the emergent nodes could provide designers with implicit feedback and opportunities to incorporate player creativity into the game narrative in an iterative design process.