跳转至

Meta$^n$：通过涌现深度实现递归自我提升

文章背景与核心概要

传统的自我提升大语言模型（LLM）智能体通常专注于精炼最终答案，而不是演进用于生成这些答案底层过程。现有的递归或自我编辑框架面临着局限性：添加元级别的系统将其保持静态，而修改自身内部代码的系统必须保留关键的编辑机制不动以维持稳定性——这从根本上将它们的元深度限制在两层左右。

本文介绍了 Meta\(^n\)，这是一个新颖的框架，它保持核心元操作（\(\Omega\)）固定，同时对输入进行完全递归。通过将 \(\Omega\) 重复应用于自身的输出——分析求解器堆栈轨迹以及生成它们的代码——Meta\(^n\) 将后续层动态构建为战略预处理器，并配有可调用的辅助库。由于基础操作 \(\Omega\) 保持不变，系统稳定性得以保存，而严格增长的输入允许每个连续层从不断提升的视角进行推理。深度是通过收敛动态决定的，而不是预先定义的，并且进化档案库会探索最佳的层链。

在两种主干架构上，Meta\(^n\) 在所有八个基准测试系列中均超越了现有的自我提升智能体。最值得注意的是，在专为抵制死记硬背技能而设计的 ARC-AGI-2 基准测试中，Meta\(^n\) 是唯一得分大于零的系统。消融研究表明，性能提升主要源于层之间传递的累积条件作用，在没有显式提示词工程的情况下，促成了独特、专门化层角色的涌现。

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

Summary

Traditional self-improving Large Language Model (LLM) agents typically focus on refining final answers rather than evolving the underlying process used to generate them. Existing recursive or self-editing frameworks face limitations: systems that add a meta-level hold it static, while those that modify their own internal code must leave critical editing machinery untouched to remain stable—inherently capping their meta-depth at around two.

This paper introduces Meta\(^n\), a novel framework that keeps the core meta-operation (\(\Omega\)) fixed while recursing entirely on its inputs. By repeatedly applying \(\Omega\) to its own outputs—analyzing the solver stack traces alongside the code that generated them—Meta\(^n\) dynamically constructs subsequent layers as strategic pre-processors paired with callable helper libraries. Because the base operation \(\Omega\) remains unchanged, system stability is preserved, while the strictly growing input allows each successive layer to reason from an increasingly elevated vantage point. Depth is determined dynamically via convergence rather than being predefined, and an evolutionary archive explores optimal layer chains.

Across two backbone architectures, Meta\(^n\) surpasses existing self-improving agents across all eight benchmark families. Most notably, on ARC-AGI-2—a benchmark specifically designed to resist memorization of rote skills—Meta\(^n\) stands alone in scoring above zero. Ablation studies reveal that the performance gains stem primarily from the cumulative conditioning passed between layers, fostering the emergent appearance of distinct, specialized layer roles without explicit prompt engineering.

Metadata & Reference Information

  • arXiv Identifier: arXiv:2608.24735 [cs.AI]
  • Authors: Zae Myung Kim, Young-Jun Lee, Seungyeon Jwa, Dongyeop Kang
  • Submitted: August 25, 2026
  • Primary Subject: Artificial Intelligence (cs.AI)
  • Secondary Subjects: Computation and Language (cs.CL); Systems and Control (eess.SY)
  • DOI: 10.48550/arXiv.2608.24735

Access & Resources