Iteration is all you need: ARC-2 solver inspired by Grover's search algorithm (github.com)

A new ARC-2 solver frames ARC-AGI-2 as an iterative search inspired by Grover’s quantum search: generate many program-synthesis attempts, execute them as an external “oracle,” grade each (0–10), and reorder the context so the highest-scoring attempts appear last—amplifying their influence via attention in the next generation. Implementation details: five main iterations, 3–5 program attempts per iteration, context saturation with all prior attempts (including failures), interleaved context-free attempts with different seeds to promote exploration (leveraging grok-4-fast’s high variance), and three static visual-analysis modules to inject productive entropy. The analogy to Grover is abstract but useful: initial attempts are a superposition, execution marks promising candidates, and grading+reordering acts like a diffusion step that concentrates attention on better solution patterns. On 60 ARC-AGI-2 sample problems the solver reached 38% success (23/60), or 53.5% when excluding Cloudflare execution errors; failures included wrong outputs, timeouts, and infra errors. Key empirical findings: detailed runtime debug info often harms search by anchoring bad paths; naive image analysis can mislead but is essential for some tasks, so mixing “untainted” attempts prevents premature fixation. The work reinforces a broader lesson across ARC solvers: iteration plus feedback (grading, execution diffs, hindsight learning) converts weak single-shot models into competitive search engines. A proposed next step is “negative context” tokens to actively suppress dead-end patterns—completing the search toolkit by enabling both attraction and repulsion in context-driven exploration.