Do LLMs Possess Strategic Intelligence? Testing LLMs in Iterated Prisoner’s Dilemmas

This study systematically evaluates the strategic capabilities of large language models through evolutionary tournaments in the Iterated Prisoner's Dilemma. We tested 28 agents, including 12 contemporary large language models from four major providers (OpenAI, Anthropic, Google, Mistral), 13 canonical game theory strategies, and 3 adaptive learning algorithms across varying interaction horizons, temperatures, and memory conditions. The experimental framework employed five-phase evolutionary tournaments with fitness-proportional selection, examining both anonymous memory and opponent tracking mechanisms to investigate how historical information impacts strategic behaviour. Results reveal that most large language models maintained high cooperation rates even in short-horizon scenarios where game theory predicts universal defection, with only GPT-5-mini demonstrating sophisticated horizon-sensitive strategic adaptation. To deepen understanding, future evaluations should include longer games, deeper tournaments, expanded cross-model comparisons, and richer environmental variations to more reliably characterise strategic behaviour across horizons and evolutionary depth.