Meta^n: Recursive Self-Improvement through Emergent Depth
The paper claims recursive meta-layers improve the process without rewriting the machinery that creates them.
Meta^n keeps a fixed operation, Ω, and repeatedly feeds it the solver traces and code from lower layers. Each new layer writes a strategic pre-process plus callable helpers, with depth determined by convergence rather than a preset limit. The authors report gains across eight benchmark families on two backbones, including the only above-zero result on ARC-AGI-2. Ablations attribute most of the recursion gain to conditioning passed between layers. HF Daily Papers' note
Meta^n keeps a fixed operation, Ω, and repeatedly feeds it the solver traces and code from lower layers. Each new layer writes a strategic pre-process plus callable helpers, with depth determined by convergence rather than a preset limit. The authors report gains across eight benchmark families on two backbones, including the only above-zero result on ARC-AGI-2. Ablations attribute most of the recursion gain to conditioning passed between layers. HF Daily Papers' note
score 5