Google DeepMind adds Werewolf and poker to Kaggle Game Arena

From perfect to imperfect information

Game Arena launched last year with chess, a game of perfect information where every piece is visible and the only limit is how far a model can reason ahead. Google DeepMind is now adding two games that deliberately withhold information from the players.

Chess is a game of perfect information. The real world is not.Montana Labs

That framing is the organizing idea of the whole announcement. Werewolf tests whether a model can extract signal from ambiguous dialogue, and poker tests whether it can act sensibly when it cannot see its opponent's cards. Both are meant to stand in for decisions made without complete data.

What the chess leaderboard already shows

The chess results give the clearest read on model progress. Gemini 3 Pro and Gemini 3 Flash hold the top Elo ratings, and Google DeepMind describes a significant jump over the Gemini 2.5 generation. The post is explicit that this is not brute-force play: unlike Stockfish, which evaluates millions of positions per second, the models rely on pattern recognition and 'intuition' to narrow the search space.

The claim that the models' internal 'thoughts' reference concepts like piece mobility, pawn structure, and king safety is the kind of detail that makes Game Arena useful as a tracking tool rather than a single-number scoreboard — it lets you watch how capability changes across model generations.

Werewolf as a deception sandbox

Werewolf is the more novel addition. Google DeepMind calls it its first team-based game played entirely through natural language, where villagers must identify hidden werewolves purely through conversation. The announcement points to a specific behavior: models spotting inconsistencies between a player's public claims and their voting patterns, then using that to build consensus with teammates.

The safety angle is the part worth flagging. Because winning requires playing both the truth-seeker and the deceiver, the benchmark doubles as a way to red-team a model's own capacity for deception in a setting with no real-world stakes. That is a deliberate design decision — testing manipulation detection and manipulation ability at the same time, inside a closed game.

Poker, and what the game selection implies for agent evaluation

Poker rounds out the set by isolating a third skill. Google DeepMind frames it not as alliance-building but as quantifying uncertainty: inferring opponents' hands and adapting to their styles in Heads-Up No-Limit Texas Hold'em. The full poker leaderboard is set to be revealed on February 4 after the tournament finals, with livestreams featuring poker players Nick Schulman, Doug Polk, and Liv Boeree, and Grandmaster Hikaru Nakamura on chess.

The specific implication is in the taxonomy Google DeepMind has now built. Chess measures long-term planning, Werewolf measures negotiation under ambiguity, and poker measures risk management — three distinct axes the post ties directly to the 'soft skills' enterprise agents need to collaborate with humans and other agents. Rather than one leaderboard for general intelligence, this is an argument for evaluating agents against separable capabilities, each with its own game.