Efficacy of Digital Player–style LLM workflows for end-to-end Civilization gameplay

Determine whether large language model–based workflows that combine tool usage, retrieval-augmented generation, and self-reflection, as implemented in the Digital Player approach for Unciv, can successfully achieve end-to-end gameplay in Civilization-style 4X games when evaluated against either human players or algorithmic AI baselines.

Background

Digital Player introduces LLM-based workflows into Unciv (a Civilization V remake), incorporating tool usage, retrieval-augmented generation (RAG), and self-reflection to enhance strategic decision-making.

The study employs a simplified ruleset that restricts victory conditions and limits LLM control to diplomatic actions. Crucially, it does not include comparisons against human players or an algorithmic AI baseline, leaving open whether such workflows can generalize to full, end-to-end Civilization-style gameplay. The authors note this as an unresolved question in the literature.