Files
myWiki/concepts/selective-hitl.md

42 lines
1.5 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: "选择性 HITL — Selective Human-in-the-Loop"
created: 2026-06-19
updated: 2026-06-19
type: concept
tags: [hitl, automation, ai-safety, decision-threshold]
sources:
- https://towardsdatascience.com/six-choices-every-ai-engineer-has-to-make-and-nobody-teaches/
---
# 选择性 HITLSelective Human-in-the-Loop
## 定义
选择性 HITL 是解决 [[human-in-the-loop|HITL]] 规模化问题的策略:不是审查每个模型决策,而是**仅在特定条件触发时引入人工审查**。
## 为什么需要选择性 HITL
完全人工审查面临三个根本问题:
1. **无法规模化**——审查量与模型吞吐量根本不匹配
2. **速度瓶颈**——实时人工干预拖慢整个系统
3. **质量退化**——审核人的不一致性降低标签质量(疲劳、偏见)
## 触发条件
仅在以下情况引入人工审查:
- **边缘案例**Edge Cases模型不确定或未见过的情况
- **低置信度输出**:预测概率低于阈值
- **高风险决策**:错误成本过高(医疗、金融、法律)
## 设计关键
1. **置信度阈值**:需要仔细校准——太低浪费人力,太高漏过错误
2. **人类推翻权限**:当人类与模型意见不一致时,是否有明确权限推翻?答案取决于领域风险
3. **反馈闭环**:被推翻的决策应回流为训练信号,持续改进模型和阈值
## 参考
- [[human-in-the-loop|Human-in-the-Loop]]
- [[ai-production-tradeoffs|AI 生产权衡]]
- [[nobrega-ai-production-tradeoffs-2026|原文文章]]