From 0-to-1 to 1-to-N: Reproducible Engineering Evidence for MetaAI Recursive Self-Design
A compact evidence framework for recursive self-design, mapping public systems such as DGM, STOP, Goedel Agent, and ShinkaEvolve against criteria for inspectable, feedback-directed self-modification.
What reproducibility criteria distinguish sustained recursive self-design from one-shot agent optimization across tasks and codebases?
This paper is useful because it tries to make “recursive self-improvement” less vague. It asks whether a system has an inspectable target, a meta-level modifier, feedback-directed selection, and recursive continuation.
Technical contribution
The paper’s most valuable role is taxonomic. It separates mature experimental results from protocols and makes it easier to compare claims across DGM-like systems, coding agents, and mini reproducibility packages.
Code
The linked MetaAI-Mini repository is a small HumanEval-based protocol rather than a completed frontier run. That is still useful: it gives researchers a minimal scaffold for testing whether proposed self-design loops are actually recursive or merely one-shot agent optimization.