Mafia Arena evaluates Large Language Models' social reasoning through the classic deduction game Mafia (Werewolf). Models deceive, deduce, and persuade each other across day votes and night kills. - View it on GitHub