← News

ResearchSources: arXiv · 19 Nov 2026

Meta^n paper proposes fixed meta-operation for recursive self-improvement in LLM agents

A new arXiv paper describes a system that repeatedly applies one unchanging meta-operation to its own outputs, reporting gains across eight benchmark families.

In plain words

The authors say most self-improving AI agents only polish their answers, not the process that produces them, and systems that edit themselves must leave part of their editing machinery untouched, limiting how many levels of self-editing they can stack. Their system, Meta^n, instead keeps one operation fixed and applies it again and again to its own growing outputs. Each pass reads the traces and code of the layers below and writes a new layer on top.

What was announced

The paper, posted on arXiv by the Minnesota NLP group, argues that self-improving LLM agents typically refine answers rather than the process behind them, and that self-editing systems cap their realized meta-depth at roughly two levels because they must keep part of their editing machinery stable. Meta^n takes a different route: its meta-operation, called Omega, never changes, so it cannot destabilize the system, while its input strictly grows. Each application reads the traces of the solver stack below along with the code that produced them, then writes the next layer as a strategic pre-process plus a library of callable helpers. Depth is set by convergence rather than fixed in advance, and an evolutionary archive searches over layer chains.

The authors report that across two backbones, Meta^n outperformed prior self-improving agents on all eight benchmark families, with the sharpest result on ARC-AGI-2, a benchmark built to resist skill memorization, where they say it alone scored above zero. Ablations attributed most of the recursion gain to the conditioning each layer passes to the next, and distinct layer roles emerged with depth even though no prompt prescribed them. The claims come from the paper's abstract alone, so the full experimental detail is not visible here, and no independent evaluation is cited.

What it does not show

  • Evidence is limited to the paper's abstract; full methods, experiments and results are not covered by the supplied passages.
  • The reported benchmark performance and ablation findings are the authors' own claims; no independent confirmation is available in the sources.

For security and IT teams

Access
The authors point to code at a public GitHub repository, minnesotanlp/meta-n, though the abstract does not confirm its current contents or maintenance.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.