Source-linked AI summary
Player-Driven Emergence in LLM-Driven Game Narrative
Xiangyu Peng, Jessica Quaye, Sudha Rao, Weijia Xu, Portia Botchway, Chris Brockett, Nebojsa Jojic, Gabriel DesGarennes, Ken Lobb, Michael Xu, Jorge Leandro, Claire Jin, Bill Dolan
TL;DR
LLM-generated stories can fall into limited patterns without human intervention, motivating the paper’s exploration of player participation in narrative creation. The authors study a mystery game with GPT-4-driven NPCs and compare player narrative graphs with the designer’s intended graph. Players created novel narrative nodes, and those creating the most tended to enjoy discovery, exploration, and experimentation.
Problem
LLM-generated stories can devolve into a limited set of patterns without human intervention, raising the question of how players might contribute to narrative design.
Source: §I · p. 1Method
The authors let players interact with GPT-4-driven NPCs in a text-adventure mystery and use GPT-4 to compare gameplay narrative graphs with the designer’s intended graph.
Source: §I · p. 1Results
Players created novel narrative nodes, and those who created the most tended to enjoy games that facilitate discovery, exploration, and experimentation.
Source: §I · p. 1, §VII · p. 8Takeaways & Limitations
The authors suggest that emergent player interactions may provide designers implicit feedback and contribute to a more collaborative model of game development.
Source: §I · p. 1, §VII · p. 7The investigation is limited to its studied setting; exploration across different LLMs, player profiles, game types, and cultural milieus is left for future research.
Source: §VII · p. 8
Abstract
from arXiv · showhide
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.
I. INTRODUCTION
The paper argues that player interaction with non-deterministic LLMs can create emergent narrative paths, while human participation may counter the limited patterns of LLM-generated stories. It investigates this idea in a text-adventure game and compares player-created narrative graphs with the designer’s intended narrative.
I. INTRODUCTION
- The paper frames player–LLM synergy as a way to produce “desire paths”: unanticipated player preferences that may be productive and beneficial to the game narrative.The authors note that LLM-generated stories can devolve into a limited set of patterns without human intervention.Source: §I · p. 1
- The authors propose involving players in the design process, with emergent interactions potentially serving as implicit feedback rather than replacing designers.They investigate this proposal in Dejaboom!, a TextWorld game with free-form NPC dialogue provided by an LLM.Source: §I · p. 1
- The authors define an emergent node as one present in the player narrative graph but absent from the designer’s original narrative graph.GPT-4 converts player and designer playthroughs into graphs for comparison.Source: §I · p. 1
- Players created new narrative nodes, including novel information-gathering strategies, suggested game elements, and alternative ways to defuse the bomb.The study involved 28 gamers; players who created the most emergent nodes tended to enjoy discovery, exploration, and experimentation.Source: §I · p. 1
II. DEJABOOM! AN LLM-DRIVEN GAME
DejaBoom! is a text-based mystery in which players repeatedly relive the day of an explosion and try to stop it. Its fixed game world combines structured objectives with dynamic, free-form NPC dialogue.
II. DEJABOOM! AN LLM-DRIVEN GAME
- Players have 30 steps per day to stop an explosion, retain memories across attempts, and encounter a reset game world and NPC memories after each explosion.Each step is one in-game action or NPC interaction.Source: A. Game Design · p. 2
- The village has six main locations, two gated secret rooms, and objects that supply disposal-kit ingredients or clues.Players can move freely between locations; the secret rooms become accessible after certain goals are achieved.Source: A. Game Design · p. 2
- Five stationary NPCs provide village information and bomb clues, combining familiar literary characters with custom characters.The design uses both kinds of personas to examine whether NPC-player interactions and player strategy differ with model-training-data representation.Source: A. Game Design · p. 2
- GPT-4 generates real-time, non-scripted NPC dialogue, with character goals and conditions that players must discover through interaction.For example, the Mad Hatter first asks for a joke before giving a riddle that clues the bomb’s location.Source: A. Game Design · p. 2
- The bomb is hidden in a secret blacksmith-shop storage room, and players must locate it and obtain a disposal kit to defuse it.The kit can be assembled from three items after consulting a recipe, or obtained from Merlin after following clues from Chef Maria.Source: A. Game Design · pp. 2–3
B. Game Implementation
The implementation layers GPT-4 over TextWorld’s fixed game logic to handle flexible player input and generate dynamic feedback and NPC responses.
B. Game Implementation
- TextWorld supplies fixed game logic, while GPT-4 makes both game input and output dynamic.In conventional text games, the command set and responses would be predefined.Source: B. Game Implementation · p. 3
- GPT-4 classifies each command as an in-game action or NPC-directed words, then routes it through the corresponding interaction path.Recognized actions are converted to a verb-object format for the fixed agent; NPC-directed words prompt an NPC response when an NPC is present.Source: B. Game Implementation · p. 3
- GPT-4 uses the command, game-agent response, and game history to generate feedback, and also receives NPC persona, backstory, clues, and reveal conditions for NPC responses.When the prompt approaches its token limit, the game history is summarized.Source: B. Game Implementation · p. 3
A. Recruitment
The study recruited experienced role-playing-game players for a one-hour play session and used pre- and post-game surveys to capture their motivations and reflections.
A. Recruitment
- The study recruited 28 U.S.-based participants with relevant role-playing-game experience; each played for an hour under observation.Participants were recruited through list-servs, received gift cards, and provided consent for use of their anonymized logs.Source: A. Recruitment · p. 3
B. Surveys
- Before playing, participants completed a motivation survey; afterward, they completed a survey reflecting on the experience’s highs and lows.Participants were given detailed game instructions before their one-hour session.Source: B. Surveys · p. 3
- Survey responses were used to categorize participants into six motivation profiles based on Quantic Foundry definitions.One profile, Creativity, describes enjoyment of games involving discovery, exploration, expression, customization, and experimentation.Source: B. Surveys · p. 4
C. Survey analysis
Most participants reported enjoying the game, especially its fluid NPC interactions, while survey responses also identified latency and response-quality issues.
C. Survey analysis
- 25 of 28 participants reported enjoyment, while 6 successfully defused the bomb.Players averaged 75 steps per hour, or 2.5 attempts per hour, with the bomb exploding every 30 steps.Source: C. Survey analysis · p. 4
- Players highlighted flexible commands and error correction, realistic NPC conversations, natural-language exploration, emergent paths, and novelty across attempts.20% mentioned flexible commands and error correction, and 20% appreciated conversations with NPCs; 10% valued natural-language movement, including all participants motivated by creativity.Source: C. Survey analysis · p. 4
- Reported issues included 15-second response latency, inconsistent NPC personas, repetitive replies, and NPCs divulging information too easily.40% reported latency, 10% persona inconsistency, 10% repetitive responses, and 14.28% wanted more mysterious NPCs.Source: C. Survey analysis · p. 4
- 54% most enjoyed fluid NPC interactions, 26% enjoyed exploring areas, and 56% wanted to play again to solve the game.Merlin and Moriarty were the most liked characters, at 53.5% and 42.86%, respectively.Source: C. Survey analysis · pp. 4–5
IV. EMERGENCE IN GAME NARRATIVE
The authors compare player gameplay graphs with a designer-intended narrative graph to identify additions to the game’s narrative. They construct player graphs by summarizing strategies from logs and merging repeated strategies across attempts.
IV. EMERGENCE IN GAME NARRATIVE
- The analysis represents each player’s strategies as nodes in a directed acyclic graph, with edges marking their temporal progression through gameplay.Player graphs are contrasted with the narrative graph derived from designers’ walkthroughs to identify emergent nodes.Source: §IV · p. 5
A. Generating Narrative Graphs
- GPT-4 segments each day’s log and summarizes each segment as a player strategy labeled with its game state.A day is one attempt; the extracted strategies form a single path through that attempt’s graph.Source: A. Generating Narrative Graphs · p. 5
- Across attempts, GPT-4 merges strategies with the same meaning and game-state label, preserving chronological order in one graph per player.The procedure successively processes each day’s log and merges it with that player’s existing graph.Source: A. Generating Narrative Graphs · p. 5
B. Identifying Emergent Nodes
The authors identify emergent nodes by comparing player narrative graphs with the designer-intended graph. The resulting additions include varied strategies and suggestions for expanding game content.
B. Identifying Emergent Nodes
- The designer-intended graph is built from multiple walkthroughs; player-graph nodes that do not merge with it are classified as emergent.The comparison uses the graph-merging procedure to distinguish unmatched player strategies from the original narrative.Source: B. Identifying Emergent Nodes · p. 5
A. Types of emergent nodes
- The 28 players produced 53 emergent nodes, 43 unique, including NPC information-gathering tactics, suggested new assets, hidden-information searches, and bomb-defusal approaches.Four nodes appeared in multiple players’ graphs.Source: A. Types of emergent nodes · p. 5
- Examples include distracting Merlin to seek his kit, searching a home for useful items, and tricking Moriarty into revealing information.The authors suggest these strategies could inform changes to prompts, game objects, or narrative branches.Source: A. Types of emergent nodes · p. 5
B. Player motivation profile-based analysis
The profile analysis reports how often players in each motivation group created emergent nodes and how many they produced. Mastery had the most creators, while the small creativity group had the highest nodes-per-creator ratio.
B. Player motivation profile-based analysis
- Across profiles, 50% to 65% of players created emergent nodes; the authors describe the profile findings as preliminary.They suggest that larger data and better profiling may help designers select nodes associated with creative profiles to tailor games to player preferences.Source: B. Player motivation profile-based… · pp. 5, 7
- Among players who declared a “creativity” profile and produced emergent nodes, two generated 11 emergent nodes in total, averaging 5.5 nodes per creator.Source: B. Player motivation profile-based… · p. 6
- For the “mastery” profile, 17 players generated 50 emergent nodes, averaging 2.94 nodes per player.Source: B. Player motivation profile-based… · p. 6
VI. RELATED WORK
Prior work has studied language models as players, game-world content generators, and participants in strategic interactions. This paper instead examines emergent behavior arising from players interacting with LLMs.
VI. RELATED WORK
- Most prior work on LLMs in interactive or text-based games studies their use for playing the game, including reinforcement-learning agents and puzzle-solving evaluations –.The review also notes a crowdsourced text-adventure platform for studying grounded dialogue and Jericho as a learning environment for text agents.Source: §VI · p. 7
- Outside text-based games, LLMs have powered Minecraft NPCs and autonomous exploration, while other work uses them to generate game content such as scripts, scenes, rooms, levels, and quests –,, [24]–.This paper instead uses LLMs for NPCs with real-time dialogue capabilities.Source: §VI · p. 7
- Prior studies examine strategic behavior emerging in LLMs, whereas this work studies emergent behavior stemming from players interacting with LLMs –[31].Source: §VI · p. 7
VII. CONCLUSION
The study finds that players interacting with a non-deterministic LLM can introduce unanticipated narrative paths and strategies. These emergent paths may be non-actionable, and the investigation is limited in scope.
VII. CONCLUSION
- Players collaborating with a non-deterministic LLM introduced unpredicted narrative paths and strategies for information gathering, game-world additions, and solving tasks.Some emergent nodes may be non-actionable or lead to dead ends, but could point toward alternative paths for richer creative play.Source: §VII · p. 7
- Players enjoyed the new capability despite frustration with latency and other aspects of the prototype game.Source: §VII · p. 7
- Players who created the most emergent nodes tended to enjoy games facilitating discovery, exploration, and experimentation.Source: §VII · p. 8
- The study leaves comparisons across LLMs, player profiles, game types, and cultural settings for future research.The authors describe the investigation's scope as inherently limited.Source: §VII · p. 8