---
id: 20260523-T0-17
title: "CP-MoE：持续学习的混合专家模型"
title_en: "CP-MoE: Catastrophic Forgetting Solution"
url: https://ai.daily.yangsir.net/daily/20260523-T0-17
issue_date: 2026-05-23
publish_date: 2026-05-22T04:00:00.000Z
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2605.20247
---

# CP-MoE：持续学习的混合专家模型

arXiv研究提出CP-MoE架构，解决大模型持续学习中的灾难性遗忘问题。该方案通过一致性保留的混合专家机制，在添加新任务时保持旧任务性能87%。测试显示，CP-MoE在10任务连续学习场景中准确率比基线高18%。这一突破将推动终身学习AI发展。

## English Version

**CP-MoE: Catastrophic Forgetting Solution**

arXiv research presents CP-MoE architecture addressing catastrophic forgetting in continual learning. The consistency-preserving mixture-of-experts mechanism maintains 87% old task performance when learning new ones. Tests show 18% higher accuracy than baselines in 10-task sequential learning scenarios, advancing lifelong AI development.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2605.20247)

**详情页**：https://ai.daily.yangsir.net/daily/20260523-T0-17

---

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