PAYADOR: A Minimalist Approach to Grounding Language Models on Structured Data for Interactive Storytelling and Role-playing Games

TL;DR

The paper tackles the world-update problem in Interactive Storytelling and RPGs, where improvised player actions can destabilize the narrative world if handled by preprogrammed rules. It introduces PAYADOR, which grounds a Large Language Model on a minimal, three-component world representation (items, locations, characters) and reframes the task as predicting post-action world-state changes using a Gemini-based predictor with constrained prompts and descriptive rendering of the world. The authors provide an open-source implementation and a playable proof-of-concept, reporting both strengths (coherence in some scenarios) and weaknesses (common-sense gaps and occasional inconsistencies) in initial experiments. By enabling consistency checks and clearer communication between automated GM and players, PAYADOR supports narrative co-creativity and offers a foundation for extensions such as automated narration and item generation, bridging classical IS/NLP approaches with modern CC methods.

Abstract

Every time an Interactive Storytelling (IS) system gets a player input, it is facing the world-update problem. Classical approaches to this problem consist in mapping that input to known preprogrammed actions, what can severely constrain the free will of the player. When the expected experience has a strong focus on improvisation, like in Role-playing Games (RPGs), this problem is critical. In this paper we present PAYADOR, a different approach that focuses on predicting the outcomes of the actions instead of representing the actions themselves. To implement this approach, we ground a Large Language Model to a minimal representation of the fictional world, obtaining promising results. We make this contribution open-source, so it can be adapted and used for other related research on unleashing the co-creativity power of RPGs.

Figures (6)

• Figure 1: A Language Model grounded on structured data for Interactive Narrative research. The roots drawing was uploaded to Wikimedia Commons by Joel Swift (user Kenizzer).
• Figure 2: An example of an error during world update, taken from an AIDungeon gameplay on April 15 2024. After the first and second user utterances, the system states that there is no bazooka in the player's inventory. However, when the player insisted, the bazooka was suddenly considered a usable weapon.
• Figure 3: A diagram summarizing the PAYADOR approach, with an example for the output of each step. It starts by rendering the World state in simple sentences "(1)", including the concatenated descriptions for each component (locations, items and characters). Using it, and the user input "(2)", a prompt is created "(3)" and the LLM is called to calculate the changes in the world "(4)". Finally, the detected changes are mapped to the appropriate instructions and the world state is updated "(5)".
• Figure 4: Testing if PAYADOR can update the world when the player wants to take two different items, and if it can prevent the player to access the kitchen without a key.
• Figure 5: Testing PAYADOR with the same sequence of actions used in Fig. \ref{['fig:AIDungeon_error']}. The occasional inconsistencies between the LLM's understanding of the world state and the actual world state become evident.