---
title: Meta^n (Recursive Self-Improvement through Emergent Depth)
type: vocabulary
url: "https://www.envisioning.com/vocab/meta-n-recursive-self-improvement"
summary: Self-improving LLM agent that keeps its meta-operation fixed and recurses on its own output, growing layer depth until convergence — achieving meta-depth well beyond the two-level cap of prior self-improvement systems.
year: 2026
generality: 0.70
---

# Meta^n (Recursive Self-Improvement through Emergent Depth)

Self-improving LLM agent that keeps its meta-operation fixed and recurses on its own output, growing layer depth until convergence — achieving meta-depth well beyond the two-level cap of prior self-improvement systems.
Meta^n (recursive self-improvement through emergent depth)

Meta^n is a recursive self-improvement framework introduced by Kim, Lee, Jwa, and Kang (University of Minnesota and Seoul National University, 2026). A fixed meta-operation Ω is applied repeatedly to its own products. Each layer reads the traces of the solver stack below together with the code that produced them, then writes the next layer as a strategic pre-process and a library of callable helpers. The meta-operation never changes, so it cannot destabilize the system. Its input strictly grows, so each layer reasons from a higher vantage than the last.

The technique addresses two structural limits of prior self-improving systems. First, systems that add a meta-level must hold that level fixed, which prevents recursion. Second, systems that edit themselves must leave part of their editing machinery untouched to stay stable, which caps meta-depth at roughly two. Meta^n removes both limits by keeping Ω fixed and recursing on its input instead of on itself. Depth is determined by convergence: the search proceeds until additional layers no longer improve validation scores. An evolutionary archive searches over layer chains.

The paper reports state-of-the-art performance on eight benchmark families across two backbones. The sharpest result is on ARC-AGI-2, which is designed to resist skill memorization. On that benchmark, Meta^n alone scores above zero while OpenEvolve and Gödel Agent score at zero. Ablations show that most of the gain from recursion comes from the conditioning each layer passes to the next. Distinct layer roles, including meta-strategist, algorithm librarian, task router, and generic remediator, emerge with depth although no prompt prescribes them. This finding links Meta^n to emergent-role-differentiation observed in multi-agent systems.

Meta^n is adjacent to recursive-self-improvement, the broader concept it instantiates, to dgm-darwin-godel-machine, a related self-modifying-agent framework, and to self-evolution, a single-agent scaffold-modification pattern. It differs by being multi-layer by construction rather than through stacked edits, and by producing emergent role differentiation as a side effect of depth.

---
Source: Envisioning — Technology Research Institute (https://www.envisioning.com/vocab/meta-n-recursive-self-improvement)