Coding agents turn toward ARC-AGI-3 — the benchmark designed to resist the pattern-matching that beat its predecessors
Recent work applies coding agents to ARC-AGI-3, the latest iteration of a benchmark built specifically to resist solutions that generalise from surface pattern statistics rather than reasoning about structure.
ARC's value has always been that it is adversarial to the field's dominant method. Each version has been redesigned after the previous one was solved in ways its authors considered unconvincing. That makes progress on it informative in a way that progress on saturated benchmarks is not.
Applying coding agents rather than raw models is the notable shift. It reframes the question from can the model see the pattern to can a system search for a program that produces it — closer to what the benchmark was arguably always testing.
Kimbodo — AI Research & Papers — July 20, 2026 → · VoltAgent — Awesome AI Agent Papers — 2026 collection →