---
id: 20260416-T0-06
title: "多智能体系统在长时任务中频繁失效，原因和位置被诊断"
title_en: "Agentic Systems Fail on Long-Horizon Tasks: Causes Identified"
url: https://ai.daily.yangsir.net/daily/20260416-T0-06
issue_date: 2026-04-16
publish_date: 2026-04-15T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2604.11978
---

# 多智能体系统在长时任务中频繁失效，原因和位置被诊断

arXiv新研究发现，大语言模型智能体在短中期任务表现良好，但在需要长时间、相互依赖行动序列的长时任务中经常失效。尽管智能体系统进展迅速，这些长时任务失败问题仍未解决。论文分析了这些失效发生的具体位置和原因，为改进多智能体系统提供了重要参考。

## English Version

**Agentic Systems Fail on Long-Horizon Tasks: Causes Identified**

New arXiv research finds that LLM agents excel at short- and mid-horizon tasks but frequently break down on long-horizon tasks requiring extended, interdependent action sequences. The study identifies where and why these agentic systems fail, providing crucial insights for improvement.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2604.11978)

**详情页**：https://ai.daily.yangsir.net/daily/20260416-T0-06

---

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