52 lines
2.8 KiB
Markdown
52 lines
2.8 KiB
Markdown
---
|
||
title: "多智能体系统可靠性"
|
||
type: source
|
||
tags: [multi-agent, reliability, architecture]
|
||
sources: [raw/AI/Multi-Agent System Reliability.md]
|
||
last_updated: 2026-04-18
|
||
---
|
||
|
||
## Source File
|
||
- [[raw/AI/Multi-Agent System Reliability.md]]
|
||
|
||
## Summary
|
||
- 核心主题:多智能体系统的可靠性架构模式
|
||
- 问题域:LLM 不可靠性导致的系统级错误传播
|
||
- 方法/机制:层级结构、共识投票、对抗式辩论、淘汰制四种架构模式
|
||
- 结论/价值:将 LLM 视为不可靠组件,通过架构设计强制正确性而非依赖模型"小心谨慎"
|
||
|
||
## Key Claims
|
||
- LLM 本质不可靠(幻觉、逻辑谬误、上下文漂移),多智能体拓扑会将错误传播到无法使用的程度
|
||
- 层级结构通过依赖图强制 Worker 协作,验证器捕获作弊
|
||
- 共识投票中 3 个模型同时幻觉相同谎言的概率仅为 0.8%(20%³)
|
||
- 对抗式辩论模拟人类"恐惧"机制,通过外部批评者纠正模型"好好先生"倾向
|
||
- 淘汰制将 LLM 视为"牲畜"而非"宠物",失败即替换而非修复
|
||
|
||
## Key Quotes
|
||
> "We don't need AI that 'cares.' We need AI that is constrained, verified, pruned, and challenged." — Alex Ewerlöf
|
||
|
||
> "Don't anthropomorphize LLMs! Find a way to piggy back on their human-corpus training while being aware of their non-biological differences." — Alex Ewerlöf
|
||
|
||
## Key Concepts
|
||
- [[多智能体系统可靠性]]:通过架构模式提升 LLM 多智能体系统可靠性的方法论
|
||
- [[层级结构 (Hierarchy)]]:Planner 分解任务 → Worker 执行 → Validator 验证的三层架构
|
||
- [[共识投票 (Consensus)]]:多数票机制抵消单个模型的随机噪声
|
||
- [[对抗式辩论 (Adversarial Debate)]]:生成器+批评者+评委的三角制衡结构
|
||
- [[淘汰制 (Knock-out)]]:适者生存的遗传算法式选择机制
|
||
- [[可靠性工程]]:将 LLM 视为分布式系统中不可靠组件的工程思维
|
||
|
||
## Key Entities
|
||
- [[Alex Ewerlöf]]:作者,27年经验的资深工程师,KTH 系统工程硕士,专注可靠性工程和弹性架构
|
||
- [[KTH]]:瑞典皇家理工学院(KTH Royal Institute of Technology)
|
||
|
||
## Connections
|
||
- [[多智能体系统可靠性]] ← extends ← [[Multi-Agent-Team]]
|
||
- [[层级结构 (Hierarchy)]] ← depends_on ← [[依赖图]]
|
||
- [[共识投票 (Consensus)]] ← uses ← [[复合 SLO]]
|
||
- [[淘汰制 (Knock-out)]] ← implements ← [[遗传算法]]
|
||
- [[对抗式辩论 (Adversarial Debate)]] ← avoids ← [[群体思维]]
|
||
- [[可靠性工程]] → applies_to → [[SRE]]
|
||
|
||
## Contradictions
|
||
- 与 [[Multi-Agent-Team]] 视角不同:Multi-Agent-Team 强调 Agent 个性和协作流程,本文强调将 LLM 视为不可靠组件的工程视角
|
||
- 与 [[Agent Chain]] 区别:Agent Chain 是简单的串联模式,本文强调验证和反馈机制 |