Elo-per-token Analysis for LLM Agent Test-Time Strategies
September 13, 2026
Elo-per-token analysis tracks the best solution found at each token budget using a Bradley-Terry model to measure how agents scale performance. This method allows for observing progress through long trajectories in open-ended tasks up to 100M tokens.
HOW THIS AFFECTS YOU
●
builderThis provides a metric to evaluate the efficiency of your agent's reasoning and tool-use trajectories.
●
researcherYou can now more accurately benchmark how agent performance scales with test-time compute.