Meta^n Paper: Recursive Self-Improvement via Emergent Depth
rohanpaul_ai · x · 2026-08-30
Researchers from the Univ of Minnesota and Seoul Univ propose Meta$n$, a novel approach for recursive self-improvement in LLM agents. Unlike traditional methods that hit a stability cap at roughly 2 meta-levels, Meta$n$ keeps the improvement operation $\Omega$ fixed and recurses on its input.
- Mechanism: It reads the code and traces from the solver stack below, writing the next layer as a strategic pre-process and library of helpers. Depth is determined by convergence rather than fixed in advance.
- Results: Meta$n$ outperforms prior self-improving agents across all 8 benchmark families on two backbones. Notably, it is the only system to score above zero on ARC-AGI-2, a benchmark designed to resist skill memorization.
- Insights: Ablation studies show that gains come from the conditioning passed between layers, and distinct roles emerge spontaneously with depth despite no explicit prompting.
Related event: Meta^n Architecture Enables Recursive Self-Improvement in AI Agents(3 posts)→
More from Research
- New 'Agentic Coding Index' Measures Coding Intelligence Density Across Models — Informal-Trouble2183 · 2026-08-31
- View: Jailbreak Discovery is More Principled Than Defense — nabla_theta · 2026-08-31
- Paper highlights GRPO gradient conflict flaw, proposes Bayesian fix — kastnerkyle · 2026-08-31
- AI and Folk Cartesianism: Analyzing philosophy in AI debates — AndyMasley · 2026-08-31
- Fields Medalist: Frontier Models Surpass Me in Many Math Tasks — littmath · 2026-08-31
- Exploring Object Detection Prompt Optimization for LLMs in DSPy — lateinteraction · 2026-08-31