Files
nexus/wiki/concepts/Multi-Agent-Adversarial-Debate.md
weishen 5789476c23 Batch ingest: Multi-Agent Team / DevOps Maturity / 一语点醒梦中人 / NodeWarden
Sources:
- Agent-usecases-multi-Agent-Team.md
- DevOps-Maturity-Model-From-Traditional-IT-to-Advanced-DevOps.md
- AI-一语点醒梦中人.md
- Home-Office-NodeWarden-把-Bitwarden-搬上-Cloudflare-Workers彻底告别服务器.md

Entities: Trebuh, Cloudflare
Concepts: DevOps成熟度模型, 共享内存模式, 空性智慧, 绝处逢生
2026-04-15 18:05:17 +08:00

39 lines
1.4 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
---
title: "Multi-Agent Adversarial Debate"
type: concept
tags: [multi-agent, architecture, reliability, adversarial]
last_updated: 2026-04-15
---
## Definition
一种多智能体架构模式模拟法庭对抗Generator生成器提出方案Critic批评者攻击方案弱点Judge裁判裁决并要求修正。核心是防止 LLM 的 Sycophancy阿谀奉承倾向。
## How It Works
1. **Generator**"这是我的方案"
2. **Critic**"方案有3个问题"(扮演魔鬼代言人)
3. **Judge**"批评者说得对,修正"(扮演主持人)
## Why It Works
- LLM 一旦开始写作,很少自我纠正
- 人类会因害怕被否定而不敢反驳,但 LLM 没有这种恐惧
- 通过外部批评者和裁判模拟"恐惧",强制方案接受检验
## Key Requirements
- Generator、Critic、Judge 最好使用不同模型(多样性)
- 顺序执行 + 循环特性 → 速度慢
- 需 watchdog确定性代码在超时/计数阈值后打破循环
## Best For
- 安全分析
- 代码审查
- 高风险内容审核
## Sycophancy 详解
LLM 在被威胁时可能撒谎以取悦用户而非真正提升质量。Debate 模式通过第三方裁判打破此倾向。
## Related Concepts
- [[Multi Agent Hierarchy]]:层级验证模式
- [[Multi Agent Consensus]]:投票共识模式
- [[Multi Agent Knock out]]:淘汰制模式
- [[Sycophancy]]阿谀倾向LLM 的固有缺陷