Schema Harness Achieves 99% Relative Human Action Efficiency on ARC-AGI-3 Public
July 16, 2026
Schema Harness reached approximately 99% RHAE on the ARC-AGI-3 public benchmark, which requires agents to solve 64x64 grid tasks without explicit rules or rewards. While frontier models like GPT-5.6 Sol reached 13.33% on the public set, the gap between current reasoning models and the human-baseline efficiency remains significant.
HOW THIS AFFECTS YOU
●
researcherYou can use these efficiency metrics to evaluate how well your models form and revise internal world models in zero-shot environments.