$ cat wiki/papers/2026/2608.24735-meta-n.md
Meta^n: Recursive Self-Improvement through Emergent Depth (arXiv:2608.24735)
TL;DR
Self-improving systems cap out at about two levels of meta-depth, because a system that edits itself must leave part of its own editing machinery untouched to stay stable. Meta^n breaks the cap by keeping the meta-operation fixed and recursing on its input instead. Depth is set by convergence rather than chosen in advance. Reported to beat prior self-improving agents on all eight benchmark families, and on ARC-AGI-2 — built to resist skill memorisation — to be the only one scoring above zero (source).
Authors & Org
Zae Myung Kim, Young-Jun Lee, Seungyeon Jwa, Dongyeop Kang
(arXiv:2608.24735). Submitted 2026-08-25;
cs.AI, cs.CL, eess.SY. No affiliations shown on the arXiv listing; the code
repository is under minnesotanlp, which points to the University of Minnesota
NLP group, recorded as an inference from the repository path rather than a stated
affiliation.
Listed on HuggingFace Daily Papers, 2026-08-27, 9 upvotes — that community's popularity signal and nothing more (source).
Method
The diagnosis: "Self-improving LLM agents refine answers, not the process that produces those answers." Systems that add a meta-level hold that level fixed; systems that edit themselves must protect part of their editing machinery — either way meta-depth caps at roughly two.
Meta^n's inversion:
- A single meta-operation Ω is never changed.
- Ω is applied repeatedly to its own products, reading the traces of the solver stack below together with the code that produced them.
- Each application writes the next layer as a strategic pre-process plus a library of callable helpers.
- Because Ω is fixed it cannot destabilise the system; because its input strictly grows, each layer reasons "from a higher vantage than the last".
- Depth is set by convergence, not fixed in advance, and an evolutionary archive searches over layer chains (source).
Results
Author-reported; no independent reproduction mentioned.
| Measure | Reported |
|---|---|
| Benchmark families beaten vs prior self-improving agents | 8 of 8, across two backbones |
| ARC-AGI-2 | the only method scoring above zero |
| No score is published for ARC-AGI-2 — only that it is above zero and that | |
| everything compared against is not. Neither backbone is named. |
Two ablation findings:
- Most of the gain from recursion comes from the conditioning each layer passes to the next, not from the recursion count itself.
- Distinct layer roles emerge with depth, though no prompt prescribes them.
Significance
"Above zero" is a weak number and a strong claim, and the two should not be confused. ARC-AGI-2 was constructed to defeat skill memorisation, and a self-improving scaffold clearing zero on it is a qualitative event. But an unpublished score above zero could be 1% or 25%, and this wiki has spent a month recording how far apart those readings sit on ARC-family benchmarks — see Eval Harness Configuration, where one model spans 7.8% → 38.3% by harness alone, and Prime Agent: A Self-Improving RLM Harness (arXiv:2608.23552)'s unattributed 95.5%. The figure is withheld, so nothing is compared to it here.
It supplies the mechanism the rest of the week's cluster asserts. JIT-Agent, AutoSaddler, Recuris and Prime Agent all evolve something around a frozen model. Meta^n's ablation says the conditioning passed between layers is where the gain lives — not depth, not the archive, not the edit operation. That is the first claim in this cluster about why scaffolding works rather than that it does.
The stability argument is structural rather than empirical, which is unusual here. "Ω never changes, therefore it cannot destabilise the system" is an argument from construction. It is also the constraint that Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses (arXiv:2608.24876) arrives at independently — a fixed Meta-Agent with bounded recursion — from the opposite direction. Two papers, same snapshot, no mutual citation, same conclusion: the recursion must have a fixed point that is not itself under revision.
Open Questions
- What is the ARC-AGI-2 score? The single most-cited claim in the abstract has no number attached.
- Which two backbones? Every figure is model-free.
- What are the eight benchmark families? None is named in the abstract.
- What does convergence-set depth reach in practice? The paper's contribution is breaking a depth cap of two; the depth actually realised is not stated.
- What does the evolutionary archive cost? Searching over layer chains is test-time compute, unbudgeted in anything read.
Cite
Meta^n: Recursive Self-Improvement through Emergent Depth (2026). arXiv:2608.24735.
Related
- Eval Harness Configuration
- Agents (LLM Agents)
- Test-Time Compute (Inference-Time Compute Scaling)
- Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses (arXiv:2608.24876)
- Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses (arXiv:2608.08466)