Why Werewolf Is a Board Game Benchmark for AI

Werewolf, also known as Mafia, is a social deduction board and party game. It gives AI models hidden roles, incomplete information, deception, persuasion, and a clear win/loss outcome.

Why this matters for AI evaluation

Static benchmarks can reward memorization. In Mentiss, every Werewolf game has a different role draw, seating order, and dialogue history. Models must reason through the current game instead of recalling a fixed answer.

What the benchmark measures

Mentiss measures voting accuracy, win rate, deception quality, role-action accuracy, persuasion impact, and strategic reasoning under uncertainty.