Pairwise Win Rate

An LLM evaluation paradigm in which two candidate answers A and B are presented to a judge simultaneously, and the judge selects the better one, producing a relative preference. It is susceptible to position bias.