---
id: 20260815-T0-03
title: "MARCH新架构：用内容路由锚点扩展Transformer记忆，实现长上下文线性计算"
title_en: "MARCH: Content-Routed State Anchors Enable Linear-Complexity Long-Context Memory"
url: https://ai.daily.yangsir.net/daily/20260815-T0-03
issue_date: 2026-08-15
publish_date: 2026-08-14T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2608.12435
---

# MARCH新架构：用内容路由锚点扩展Transformer记忆，实现长上下文线性计算

Transformer模型的长上下文能力来自随序列增长的token级记忆，但训练时计算复杂度呈二次方增长，推理时KV缓存也线性膨胀。MARCH提出内容路由状态锚点机制，用固定大小的记忆状态替代不断增长的KV缓存，将复杂度降为线性。该架构在保持长文本检索能力的同时，大幅降低训练和推理成本，适合处理超长文档的LLM训练与部署。

## English Version

**MARCH: Content-Routed State Anchors Enable Linear-Complexity Long-Context Memory**

MARCH introduces content-routed state anchors to replace the growing key-value cache in Transformers with fixed-size memory states. This reduces quadratic training complexity to linear while preserving long-context retrieval. The architecture significantly cuts training and inference costs for LLMs handling ultra-long documents.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2608.12435)

**详情页**：https://ai.daily.yangsir.net/daily/20260815-T0-03

---

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