LLMs and Rule-Based Systems Combined for Interactive Storytelling
A new study from arXiv (2605.24719) explores combining Large Language Models (LLMs) with rule-based systems to improve story coherence in interactive storytelling. Researchers tested Llama 3 70B and Gemini 1.5 Flash in English and Spanish, with eight participants playing two scenarios. The neuro-symbolic architecture aims to leverage LLMs for predicting state changes while using pre-programmed world-state transformations to maintain narrative consistency, addressing incoherence issues common in purely LLM-based approaches.
Key facts
- arXiv paper 2605.24719 investigates neuro-symbolic interactive storytelling
- Uses Llama 3 70B (open-source) and Gemini 1.5 Flash (closed-source) models
- Testing conducted in English and Spanish
- Eight participants played two scenarios
- Aims to address story coherence problems in LLM-only systems
- Combines LLMs with rule-based world-state transformations
- Exploratory evaluation of player expression and coherence
- Published on arXiv
Entities
Institutions
- arXiv