# STORY2GAME: Generating (Almost) Everything in an Interactive Fiction Game

**Authors:** Eric Zhou, Shreyas Basavatia, Moontashir Siam, Zexin Chen, Mark O. Riedl  
**arXiv:** [2505.03547](https://arxiv.org/abs/2505.03547)  
**Date:** 2025-05-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2505.03547v1/

## Key points

- Introduction of STORY2GAME: a new approach using Large Language Models for text-based interactive fiction games
- Process involves crafting narrative, constructing game world, and coding actions in a game engine for interactive storytelling
- Emphasis on open-ended story generation while maintaining connection to game state
- Success rates of action code generation analyzed, focusing on successful actions and seamless execution in stories
- Challenges with object misidentification by LLM due to varying adjectives used for objects in descriptions
- Exploration of dynamic action generation by prompting LLM to provide unique verbs for items and characters
- New items commonly used as preconditions for actions to enhance player interactions within the game
- Examination of how different types of preconditions impact action generation for characters and items, highlighting trends based on interaction nature
- Potential of Large Language Models in creating immersive story-driven gaming experiences with dynamic action capabilities

## Summary

Our research introduces STORY2GAME, a groundbreaking approach that leverages Large Language Models to generate text-based interactive fiction games. This innovative method begins by crafting a narrative and constructing the game world, followed by coding actions in a game engine to enable interactive storytelling. Unlike traditional approaches that rely on predetermined actions, our method allows for more open-ended story generation while ensuring experiences remain rooted in the game state. We delve into the success rates of action code generation and highlight the percentage of successful actions and the number of stories where all actions were executed seamlessly. Our findings show that most sentences compile successfully, with occasional challenges arising from object misidentification by the LLM. This issue typically occurs when objects are described with varying adjectives, leading to ambiguity in their identification within the game state. Furthermore, we explore dynamic action generation and evaluate the completeness and coherence of newly generated actions. By prompting the LLM to provide verbs for items and characters in a story, we ensure that these verbs are unique and enhance player interactions within the game. Our analysis reveals that new items are commonly used as preconditions for actions due to their integral role in gameplay mechanics. Additionally, we examine how different types of preconditions—such as creating new objects or attributes—impact action generation for characters and items in a story. We observe varying trends in precondition selection based on the nature of interactions with characters versus objects, shedding light on how novel actions can be executed within the game environment. Overall, our research showcases the potential of using Large Language Models for generating interactive fiction games with dynamic action capabilities. By refining our approach to address challenges like object misidentification and optimizing precondition selection for diverse gameplay scenarios, we aim to enhance player engagement and create immersive story-driven gaming experiences.