Poker Arena: Multi-Axis Profiling of Strategic Reasoning and Memory in LLMs

Poker Arena introduces a novel platform for evaluating strategic reasoning in large language models (LLMs) through a no-limit Texas Hold'em tournament setup, utilizing a three-layer memory architecture that captures within-hand, session, and cross-session dynamics. The study assesses seven LLMs over 50 sessions, revealing that performance metrics based on tournament chips and cognitive axis scores can produce divergent rankings, highlighting the importance of multi-axis profiling in understanding model capabilities. This approach emphasizes that consistent performance across various reasoning dimensions is more indicative of practical effectiveness than peak performance in isolated areas, which is crucial for AI practitioners aiming to enhance strategic decision-making in uncertain environments.