Continual-ARCby unify

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

Program library

Every answer is a program, kept once it reproduces the demonstrations and run first when the task returns.

Held-out tasks, budget used
42.8 solved 69%
Model spend
$0.53 767 calls
Instances
100 2.5 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