Multi-Agent Strategic Games with LLMs

A new paper from arXiv (cs.GT) by Maxim Chupilkin explores whether large language models can be used to study the strategic foundations of conflict and cooperation. The author uses LLMs as experimental subjects in a repeated security dilemma — a classic game theory scenario modeling international relations. The baseline game is extended along three theoretically central dimensions: multipolarity (more than two players), finite time horizons (known end to the game), and the availability of communication between agents. Across multiple LLMs, results exhibit systematic patterns: multipolarity increases conflict likelihood, finite horizons induce universal unraveling consistent with backward-induction logic (players defect earlier as the end approaches), and communication reduces conflict by enabling signaling and reciprocity. Beyond observed behavior, the design provides access to agents' private reasoning and public messages, linking choices to underlying strategic logics like preemption, cooperation under uncertainty, and trust-building.

The contribution is primarily methodological: LLM-based experiments offer a scalable, transparent, and replicable approach to probing theoretical mechanisms in game theory and international relations. This opens the door for researchers to test political science hypotheses without human subjects, while gaining insight into the reasoning paths that lead to cooperation or conflict. The findings also have practical implications for AI alignment and multi-agent system design, suggesting that LLMs can simulate complex strategic interactions and that their behavior is sensitive to game structure in theoretically predictable ways.