Definition
Plain language
A tournament-style way to pick the best of many AI attempts at a task by comparing them in pairs.
As stated in the literature
Recursive Tournament Voting, a test-time scaling method that uses an LLM judge to compare structured summaries of rollouts in a single-elimination bracket.
Also called: Recursive Tournament Voting
Why it matters: Pairwise tournaments scale gracefully to many candidates and play to judge models' strength at comparison rather than absolute scoring, often yielding better picks than naive top-k.
For example, sixteen candidate answers are paired up in a bracket; the judge picks a winner from each pair, then winners face off, until one finalist remains.
Heard on the show
“They call it Recursive Tournament Voting, RTV for short.”Episode 003 — How to Pick the Best of Sixteen Coding Agent Rollouts