LMArena overhauls agent leaderboard: Agent Arena Index scores 2.9M sessions with anchored 1000-point scale

arena · x · 2026-11-19

LMArena refreshed its agent leaderboard into the Agent Arena Index, built on 2.9M+ user sessions. It adds domain-specific indexes (Agentic Work, Agentic Coding, Agentic Chat, Task Outcome, Steering, Tool Use) while keeping the Alignment Index. Methodology: net improvement scores τm = E[Y|M=m] − E[Y] over z-scored signals (steerability, praise vs complaint), estimated with a self-normalized inverse-probability-weighted estimator; new scores are a linear transform τ̂ = 1000·τ̂m + Δ, where the anchor Δ is recomputed each release so scores of stable, well-sampled models stay comparable across releases (one index point = 1/1000 of a signal standard deviation; July average initialized to 1000).

Original post →

More from Models

Models channel →