Researchers propose Meta^n, a system where a fixed meta-operation is recursively applied to its own outputs, with each layer reading prior solver traces and producing strategic preprocessing plus helper libraries for the next layer. Because the operation itself never changes, the approach enables unbounded meta-depth without the instability seen in prior self-improving agents that refine answers or fix meta-levels. Across multiple backbones, Meta^n surpasses earlier self-improving agents on eight benchmark families, notably achieving the only non-zero performance on ARC-AGI-2.
