Continual-ARCby unify

ReplayARC-AGI-1 · Fresh · GPT-6 Luna, high

Demo-cache floor

The model alone, shown every demonstration it bought for the task again for free.

Held-out tasks, budget used
26.3 solved 92%
Model spend
$1.08 477 calls
Instances
100 6 hours

Budget used is the share of the cap of 16 that an instance was charged, averaged over the instances: a solved instance costs 1 for each wrong answer and each demonstration, an unsolved one all 16, so lower is better.

Loading the run