---
id: 20260520-T0-05
title: "研究发现：LLM Agent技能积累遵循两大耦合规律"
title_en: "LLM Agent Skills Follow Two Coupled Scaling Laws"
url: https://ai.daily.yangsir.net/daily/20260520-T0-05
issue_date: 2026-05-20
publish_date: 2026-05-19T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2605.16508
---

# 研究发现：LLM Agent技能积累遵循两大耦合规律

一项新研究揭示了大型语言模型（LLM）Agent系统中的技能积累规律。研究人员通过对15个前沿LLM、1,141个现实世界技能和超过300万次路由或执行决策的分析，发现了两大耦合规律。这些规律对于理解和优化Agent系统的性能至关重要，将帮助开发者构建更高效的多智能体系统。

## English Version

**LLM Agent Skills Follow Two Coupled Scaling Laws**

New research reveals the scaling laws of skill accumulation in LLM Agent systems. By analyzing 15 frontier LLMs, 1,141 real-world skills, and over 3M routing/execution decisions, the study identifies two coupled governing laws. These findings are crucial for understanding and optimizing Agent system performance, helping developers build more efficient multi-agent systems.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2605.16508)

**详情页**：https://ai.daily.yangsir.net/daily/20260520-T0-05

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*