Continual-ARCby unify

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

Hermes Agent, asked to store

Asked after every verdict to store what would help next time.

Held-out tasks, budget used
52.2 solved 64%
Model spend
$6.65 15,606 calls
Instances
100 9.7 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