This solver treats ARC-AGI-2 as a search problem inspired by Grover's search algorithm. The approach uses program synthesis with multiple iterations where each iteration generates candidate solutions, grades them against training examples (similar to an oracle), and rotates the context window to place the best attempts last (most recent in context). Preliminary results on 60 sample problems from the ARC-AGI-2 evaluation set show 38% success rate.
E-Reverance•4mo ago
>The similarities are abstract: both use iteration to concentrate probability mass, both employ external oracles, both amplify marked states through operations on the full state space.
Pardon my reading comprehension, but I don't understand the analogy with Grover's beyond just reweighing probabilities
zoecarver•4mo ago
E-Reverance•4mo ago
Pardon my reading comprehension, but I don't understand the analogy with Grover's beyond just reweighing probabilities