arXiv cs.AI· Zae Myung Kim, Young-Jun Lee, Seungyeon Jwa, Dongyeop Kang·· 10 小时前AI 评分58
Meta$^n$:通过涌现深度实现 LLM 智能体的递归自我改进
Meta$^n$: Recursive Self-Improvement through Emergent Depth
AI 导读
论文提出 Meta$^n$,保持元操作 Ω 固定、对其自身产物反复递归,让每一层从更高视角推理,深度由收敛决定而非预先设定。在两个骨干模型上,Meta$^n$ 在全部八个基准族上优于此前的自我改进智能体,在 ARC-AGI-2 上是唯一得分高于零的方法。
正文 · 原文
Abstract:Self-improving LLM agents refine answers, not the process that produces those answers. Systems that add a meta-level hold that level fixed, and those that edit themselves must leave part of their own editing machinery untouched to stay stable, capping the meta-depth they realize at roughly two. We present Meta$^n$, which keeps the meta-operation fixed and recurses on its input instead. That operation, $\Omega$, is applied repeatedly to its own products, reading the traces of the solver stack below together with the code that produced them, then writing the next layer as a strategic pre-process and a library of callable helpers. Because $\Omega$ never changes, it cannot destabilize the system, and because its input strictly grows, each layer reasons from a higher vantage than the last. Depth is set by convergence rather than fixed in advance, and an evolutionary archive searches over layer chains. Across two backbones, Meta$^n$ outperforms prior self-improving agents on all eight benchmark families. The sharpest case is ARC-AGI-2, built to resist skill memorization, where it alone scores above zero. Ablations indicate that most of the gain from recursion comes from the conditioning each layer passes to the next, and distinct layer roles emerge with depth although no prompt prescribes them. Code available at this https URL
| Comments: | Accepted to NeurIPS 2026 |
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Systems and Control (eess.SY) |
| Cite as: | arXiv:2608.24735 [cs.AI] |
| (or arXiv:2608.24735v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2608.24735 arXiv-issued DOI via DataCite |
Submission history
From: Zae Myung Kim [view email]
[v1]
Tue, 25 Aug 2026 15:44:25 UTC (495 KB)
[v2]
Wed, 18 Nov 2026 03:05:24 UTC (504 KB)
来源:arXiv cs.AI · arxiv.org