# Open-Ended Wargames with Large Language Models

**Authors:** Daniel P. Hogan, Andrea Brennen  
**arXiv:** [2404.11446](https://arxiv.org/abs/2404.11446)  
**Date:** 2024-04-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2404.11446v1/

## Key points

- Snow Globe is a cutting-edge multi-agent system powered by large language models (LLMs) that automates qualitative wargames.
- The system utilizes a "history" object to record the progression of the game, with each entry associated with an agent.
- Information asymmetry is introduced by providing each player or team with their own history object.
- Players respond to situations based on history, persona description, and specific queries tailored to their roles.
- Teams operate similarly to individual players but require a designated team leader for consolidating responses into a joint action plan.
- The control acts as the moderator, adjudicating adversarial wargames by soliciting proposed actions from players and generating plausible outcomes based on their responses.
- Snow Globe simulates tabletop exercises like AI incident response scenarios and geopolitical crises, offering a versatile platform for exploring decision-making contexts.
- Its software architecture allows for flexible gameplay options involving AI, humans, or a combination of both in scenario preparation, gameplay, and post-game analysis.

## Summary

Snow Globe is a cutting-edge multi-agent system powered by large language models (LLMs) that automates qualitative wargames to enhance understanding and rehearsal of real-world decision-making processes. The system utilizes a "history" object to record the progression of the game, with each entry associated with an agent. To introduce information asymmetry, each player or team is provided with their own history object. Players are prompted to respond to situations based on the history, their persona description, and specific queries tailored to their roles. In Snow Globe, teams operate similarly to individual players but require a designated team leader to consolidate responses into a joint action plan. The control acts as the moderator, adjudicating adversarial wargames by soliciting proposed actions from players and generating plausible outcomes based on their responses. By simulating tabletop exercises such as AI incident response scenarios and geopolitical crises, Snow Globe offers a versatile platform for exploring various decision-making contexts. Its software architecture allows for flexible gameplay options where AI, humans, or a combination of both can participate in scenario preparation, gameplay, and post-game analysis. With an open-source implementation available alongside this publication, Snow Globe presents a novel approach to automating qualitative wargames and expanding the possibilities for strategic decision-making simulations within the broader wargaming ecosystem.