Why it matters

Arena is dominant preference benchmark. Understanding shapes model choice.

Advertisement

The architecture

User prompts + 2 models blind.

User picks winner.

Elo rating updates.

Chatbot Arena flowPrompt + 2 modelsblind testUser votes winnerpreferenceElo ratingleaderboardchat.lmsys.org; widely-cited but subject to gaming; category-specific arenas added
LMSYS Arena.
Advertisement

How it works end to end

chat.lmsys.org.

Blind pairwise.

Category-specific arenas (coding, hard prompts).

Style bias critique addressed with Style Control.