Tournament sampling begins with multiple independently drawn candidates. They are paired and compared across rounds until one remains, allowing a selection rule to shape the final choice without assigning a single direct score adjustment to every candidate.
In keyed text watermarking, the comparison rule prefers candidates according to secret green-list decisions. Because the initial candidates come from the original model distribution, the method can preserve marginal probabilities more faithfully than a fixed logit bias while still accumulating a detectable keyed pattern.
ELI5
Tournament sampling chooses an answer by letting several possible choices compete. The choices are compared in pairs, winners move to the next round, and one final choice remains.
For example, imagine drawing eight possible next words and holding a small knockout contest between them. The comparison rule decides each winner, which can guide the final text while still starting from words the model would normally consider.
Why use tournament sampling in text watermarking?
It can introduce a keyed preference while drawing candidates honestly from the model's original next-token distribution.
Does tournament sampling always preserve the original distribution?
Preservation depends on the tournament design and keyed selection rule, so the method needs mathematical analysis and empirical validation.

