Tournament sampling is how SynthID-Text turns a secret key into a watermark. At each position the sampler draws a set of candidate tokens from the model’s distribution, assigns every candidate a keyed pseudo-random bit (a “g-value” computed from the key, the recent context and the token), and runs a bracket: in each round, candidates are paired and the one whose bit is 1 advances. The final winner is emitted.
Because all candidates came from the model’s own distribution, the output stays fluent; the tournament only settles ties. More rounds make the winning token’s bits lean further from fair, giving a stronger detection signal at the cost of a slightly more selective distribution. The detector recomputes the bits and checks whether 1s are over-represented.