RHAE
The scoring metric behind ARC-AGI-3, which measures not just whether an agent solved a level but how many moves it needed compared with a human.
Most benchmarks ask a simple question: did the model get it right? RHAE, short for Relative Human Action Efficiency, adds a second one: how much flailing did it take? It compares the number of actions an agent spent on a level against what a first-time human player needed, then aggregates that across every level and environment in the test.
The reason this matters is that brute force is cheap for a machine. An agent can win almost any small game by trying thousands of things at random, and a pass-or-fail score would happily call that a success. RHAE does not, which makes it a better proxy for the thing people actually want from an agent: getting somewhere sensible without burning an afternoon of compute on dead ends.