Definition
Plain language
A guessing game with no stable winning move, where the only good strategy is to be unpredictable.
As stated in the literature
A zero-sum two-player game with no pure-strategy Nash equilibrium; the optimal play is a fifty-fifty mixed strategy, making it a test of whether a model can deliberately randomize.
Why it matters: It tests whether a model can be genuinely unpredictable, a skill needed wherever any detectable pattern can be used against it.
For example, in a game where you win by matching your opponent's coin flip and they win by mismatching, any predictable habit gets exploited.
Heard on the show
“Prisoner's Dilemma, Battle of the Sexes, Stag Hunt, Matching Pennies.”Episode 018 — Language Models Compute the Rational Move, Then Override It