The arena finds the ranking
From Chapter 43 of Large Language Models from the Ground Up. In a Chatbot-Arena-style contest, two anonymous models answer, a human votes, and the names are revealed. Each vote nudges an Elo rating. No single voter grades more than a handful of duels, yet a stable ranking emerges. Here the models have hidden true strengths — watch the public ratings hunt them down.
Stream anonymous battles, watch Elo converge
Each battle picks a random pair. The winner is decided by the models' hidden true strengths through the Elo expected-score formula, then both public ratings update with K = 32. Early on the leaderboard is jumpy and often wrong; after a few hundred votes it locks onto the true order. Flip Reveal true strength to see the target.
Press Play to begin the tournament.
The outcome of every duel is set by the expected-score curve E = 1/(1 + 10(Rb−Ra)/400) applied to the hidden strengths — a 400-point gap means roughly 10-to-1 odds. Ratings update by R ← R + K×(S − E) with K = 32, so an upset swings a rating hard while an expected win barely moves it (the chapter's worked case: 1200 beating 1000 gains only +7.7, but losing costs −24.3). Because every duel is zero-sum, the total of all ratings never changes. At first the order is noise; after hundreds of votes the public ratings sort into the true order — exactly how millions of anonymous LMArena votes turn into a leaderboard, even though no voter saw more than a few battles. (Real leaderboards fit all votes at once with Bradley–Terry; sequential Elo here is the teaching version.)