---
id: 20260523-T0-16
title: "ACC：代理轨迹编译实现长上下文训练"
title_en: "ACC: Compiling Agent Trajectories"
url: https://ai.daily.yangsir.net/daily/20260523-T0-16
issue_date: 2026-05-23
publish_date: 2026-05-22T04:00:00.000Z
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2605.21850
---

# ACC：代理轨迹编译实现长上下文训练

arXiv论文提出ACC方法，通过编译代理轨迹实现长上下文训练。该方案无需昂贵长文档整理，利用代理产生的自然轨迹数据。实验显示，ACC训练的模型在长文本任务上性能提升25%，训练成本降低40%。这一方法将降低长上下文AI模型的训练门槛。

## English Version

**ACC: Compiling Agent Trajectories**

arXiv paper introduces ACC method, compiling agent trajectories for long-context training. It eliminates costly long-document curation by using natural agent trajectory data. Experiments show 25% performance improvement on long-text tasks with 40% lower training costs. This lowers the barrier for long-context AI model training.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2605.21850)

**详情页**：https://ai.daily.yangsir.net/daily/20260523-T0-16

---

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