---
id: 20261107-T0-02
title: "Octopus：让多智能体共享一份潜在KV，减少上下文重复"
title_en: "Octopus: Shared Latent KV Cache for Multi-Agent LLM Systems"
url: https://ai.daily.yangsir.net/daily/20261107-T0-02
issue_date: 2026-11-07
publish_date: 2026-11-06T05:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2611.06956
---

# Octopus：让多智能体共享一份潜在KV，减少上下文重复

论文提出Octopus架构，针对多智能体LLM系统的一个痛点：现有框架通过派生隔离子智能体并行探索，但会重复预填充相同的上下文。Octopus的做法是让多个分支共享同一份潜在KV表示，即一个大脑、多个分支。这减少了重复计算，同时保留并行搜索与验证的结构。如果效果成立，多智能体系统的推理成本会随分支数增加而更平缓地上升。

## English Version

**Octopus: Shared Latent KV Cache for Multi-Agent LLM Systems**

A new paper introduces Octopus, an architecture for multi-agent LLM systems. Existing frameworks spawn isolated sub-agents to explore in parallel, but duplicate the prefill of identical context across them. Octopus has multiple branches share a single latent KV representation, one brain with many branches, cutting redundant computation while preserving parallel search and verification. If the gains hold, inference cost for multi-agent systems would scale more gently as branches multiply.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2611.06956)

**详情页**：https://ai.daily.yangsir.net/daily/20261107-T0-02

---

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