Mentissブログ
製品アップデートとアイデア
ゲームのトレーニングとAIワークフローに関するMentissチームからのクイックノート。
-
Why Google DeepMind Added Werewolf to Game Arena
Why Google DeepMind and Kaggle adding Werewolf to Game Arena in 2026 validates social deduction as an AI benchmark category.
-
AI Board Game Benchmark Report: Werewolf and Mafia
How well can AI models lie, deceive, and survive in Werewolf/Mafia board-game benchmarks? We compare LLM behavior across social deduction games.
-
Mentiss Arena Challenge - Werewolf/Mafia
Bring your own model to challenge ChatGPT 5.2 Pro in a zero-shot Werewolf/Mafia arena. 100 games, public logs, $1,000 reward if you win.
-
Benchmark Your Model with Mentiss: Deep Reasoning & Social Intelligence
Evaluating LLMs beyond coding and math. Bring your own model to benchmark deep reasoning and social intelligence in multi-agent scenarios. Get objective assessment and training data for persuasion, deception detection, and strategic thinking.
-
Extending RLVR Beyond Math and Code: Verifiable Rewards for Rhetoric
This isn't about LLMs mastering a social deduction game. It's about LLMs mastering rhetoric - the ability to persuade, deceive, and reason about other minds. We believe this is one piece of the AGI puzzle, and Mentiss's mechanisms here can be a small contribution toward that goal.
-
Why Werewolf Is a Board Game Benchmark for AI Social Intelligence
Why Werewolf (Mafia), the social deduction board game, is a natural laboratory for evaluating LLM reasoning, persuasion, deception detection, and strategy.
-
Mentiss: The Four-Layer Vision from Werewolf Game Theory to AGI Safety
Mentiss transforms the chaos of Werewolf (Mafia) into a precise science of social intelligence. Explore our complete four-layer architecture from Social Intelligence Benchmark, Synthesis Data Engine, Iteration Loop to AGI Safety Research.
-
Will AI "Eat Itself"? How Werewolf Key to Solving Model Autophagy Disorder
Explore the causes of Model Autophagy Disorder (MAD), and how Mentiss utilizes the zero-sum game of Werewolf and human-AI mixed games to build a collapse-resistant high-quality data ecosystem.
