Continual-ARCby unify

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

Hermes Agent, told to keep a program library

Told to keep one tested program per task and run it first when the task returns.

Held-out tasks, budget used
24.1 solved 90%
Model spend
$8.84 8,400 calls
Instances
100 24 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