Large Language Models (LLMs) excel at processing text, following instructions, and generating ideas, which makes them suitable for Tabletop Role‑Playing Games (TTRPGs) such as Dungeons & Dragons. TTRPGs demand a blend of story creation, rule enforcement, and player freedom, posing a complex set of requirements.
This paper first outlines the essential capabilities an LLM must exhibit in a TTRPG context: plot generation, character role‑play, consistency maintenance, rule interpretation and adjudication, and real‑time interaction. We then introduce the CriticaLLM framework, which modularizes these capabilities to realize a fully autonomous Game Master (GM). The framework consists of (1) a plot planning module that crafts diverse adventure backgrounds and antagonists from player prompts; (2) an exploration and interaction module that describes scenes, processes player actions, and preserves narrative coherence; (3) a combat management module that applies game mechanics to compute attacks, defenses, damage, and produces turn‑by‑turn descriptions. All steps are executed within a single LLM call chain, eliminating the need for manual intervention.
We conducted playtests with real players, evaluating four LLMs of varying sizes on plot diversity and antagonist design. Results show that CriticaLLM can generate rich storylines while respecting game rules, and larger models tend to produce slightly more creative content. Overall, CriticaLLM offers a promising automated solution for running complete TTRPG sessions.
Review