PAYADOR

TL;DR

PAYADOR explores how LLMs can be grounded on structured data to increase the consistency of the fictional world-state after a user input. The system models the fictional world as components (items, characters, locations, and puzzles) and uses LLMs to predict (already pre-programmed) world-state Transformations that modify that state.

How it Works

Transformations

LLMs predict Transformations—changes to the symbolic state of the world that are applied to the structured world model.

World Model

Items, Puzzles, Locations, and Characters model the fictional world state.

Consistency Checks

Structured world data enables consistency checks against basic integrity constraints (e.g., preventing movement to unreachable locations), preventing logically invalid LLM-suggested transformations.

Free-text Input

The system processes free-text player input and maps it to transformations applied to the structured world-state model.

Language Extensibility

Supports Spanish and English. The modular architecture makes extending to additional languages straightforward with minimal effort.

Publications

PAYADOR was first proposed in the ICCC'24 paper and further developed in Santiago Góngora's Master's thesis.

Featured In

Getting Started

For detailed documentation, visit the repository on GitHub.