On February 3, Google DeepMind announced the addition of two new games, Werewolf and Poker, to the Kaggle Game Arena. The platform was launched last year in collaboration with Kaggle to measure reasoning and strategic planning capabilities through chess. However, with this expansion, the company has established an environment to benchmark social reasoning and calculated risk assessment, driven by the awareness that "while chess is a perfect information game, the real world is not."
A Product Manager at the company stated that the policy is to utilize these games as controlled sandboxes to evaluate whether AI models can act safely in complex environments.
Sources: Advancing AI Benchmarking with Game Arena (HN 134pt, 54 comments) (HN Search (backfill))