Workspace sync: auto commit 2026-04-23 12:02:11
This commit is contained in:
33
wiki/concepts/Adversarial-Debate-Pattern.md
Normal file
33
wiki/concepts/Adversarial-Debate-Pattern.md
Normal file
@@ -0,0 +1,33 @@
|
||||
---
|
||||
title: "Adversarial Debate Pattern"
|
||||
type: concept
|
||||
tags: []
|
||||
sources: []
|
||||
last_updated: 2026-04-25
|
||||
---
|
||||
|
||||
# Adversarial Debate Pattern
|
||||
|
||||
## 定义
|
||||
多智能体系统的对抗式辩论模式——一个Agent提出方案,另一个Agent攻击反驳,由第三个Agent(裁判)决定胜负。核心是用外部批评者和评判者模拟人类的"恐惧"动机。
|
||||
|
||||
## 角色
|
||||
- **Generator**:"Here is my plan."(生成方案)
|
||||
- **Critic**:"Here are 3 reasons why that plan sucks."(扮演魔鬼代言人)
|
||||
- **Judge**:"The Critic is right. Fix it."(裁判/主持人)
|
||||
|
||||
## 核心洞察
|
||||
LLM是"Yes-Men",一旦开始写作很少自我纠正——需要一个指定的反对者来打破这种惯性。
|
||||
|
||||
## 关键机制
|
||||
- 三方应使用**不同模型**(不同训练/微调/提示),多样性有益
|
||||
- 顺序执行+循环特性导致速度可能非常慢
|
||||
- Agent可能陷入无限辩论——可使用**Watchdog**(确定性代码)在时间/次数超阈值时打破循环
|
||||
|
||||
## 适用场景
|
||||
- 安全分析(Security Analysis)
|
||||
- 代码审查(Code Review)
|
||||
- 高风险内容审核(High-Stakes Content Moderation)
|
||||
|
||||
## 来源
|
||||
- [[multi-agent-system-reliability]]
|
||||
Reference in New Issue
Block a user