Shall We Play a Game? Language Models for Open-ended Wargames

Abstract: Wargames are multi-faceted, multi-player depictions of conflict in which participants' decisions influence future events. Wargames are often used to explore the strategic implications of decision-making. However, it also encompasses entertainment-oriented simulations, ranging from Chess to tabletop role-playing games like Dungeons & Dragons (D&D). On the more open-ended side of the spectrum of wargames, players use natural language to convey their moves, and adjudicators propose outcomes. LLMs (LMs) are increasingly being considered for how they can provide insights into real-world, consequential decisions. We conduct a scoping literature review of a curated selection of 100 recent works on AI in wargames, from which we construct an ontology of wargames in terms of the creativity afforded to either the players or adjudicators. Focusing on the space of wargames with the most open-endedness for players and adjudicators, we distill a set of considerations for when and how to use LMs in different application areas. We also present a set of safety considerations, best practices for deploying LMs in open-ended wargames, and conclude with a set of high-impact open research challenges.