WOPR tests AI command-and-control in 'Nuclear War' wargame
TL;DR
- WOPR is a social-simulation environment that uses the published card game Nuclear War as its first test case for high-stakes organizational decisions.
- Each faction is modeled as a collective command-and-control system rather than a single agent, with a four-rung ladder from silence to private negotiation.
- Code, example configurations and replay data are published openly, with Concordia adopted as the default agent harness driving the game.
The test environment is a card game. WOPR, a social-simulation harness described in a new arxiv preprint, uses the published card game Nuclear War as its first instantiation, running agents through a "deterministic, replay-validated rules engine" in which "every strategic choice" is "an explicit agent decision."
Authors Glenn Matlin, Isaac Song, Anthony Wen-Ming Zang and Mark Riedl model each player not as a single agent but as "a collective command-and-control system," layered on top of a "four-rung press ladder from silence to private single-recipient channels with structured commitments." Concordia is the default harness driving the game.
"We start with military decision-making because of its safety implications and because it needs further study," the authors write, "but the design is not specific to it."
Code, example configurations and replay data are published on GitHub. The abstract reports no outcome statistics from runs on the engine, and does not say which agent models were tested inside the harness.
Shared on Bluesky by 1 AI expert
Originally reported by arxiv.org
Read the original article →Original headline: No One Wins in Nuclear War: A Social Simulation of Military Decision-making