The paper presents Kaggle Game Arena, an open platform for evaluating through competitive games. It includes three pilot environments: Chess (perfect information), Poker (imperfect information), and Werewolf (multiplayer). The platform allows models to play head-to-head, with performance measured by objective game outcomes like wins and losses.
By using dynamic games rather than static , the platform resists performance saturation and data contamination. The authors argue that games provide a more robust and evolving evaluation that can keep pace with model improvements.
The results from running full competitions across models are detailed in the report, but are not included in the provided text. The platform aims to be reproducible, transparent, and generalizable to new games over time.