20260601
This commit is contained in:
39
concepts/context-state-estimation.md
Normal file
39
concepts/context-state-estimation.md
Normal file
@@ -0,0 +1,39 @@
|
||||
---
|
||||
title: "Context as State Estimation(上下文作为状态估计)"
|
||||
created: 2026-05-23
|
||||
updated: 2026-05-23
|
||||
type: concept
|
||||
tags: [agent, context, state-estimation, memory]
|
||||
sources: [raw/papers/agent-harness-engineering-survey-2026.md]
|
||||
confidence: medium
|
||||
---
|
||||
|
||||
# Context as State Estimation
|
||||
|
||||
> 将上下文管理重新概念化为**状态估计问题**:Agent 的内部工作状态是对真实任务状态的估计,每次压缩、检索或遗忘操作都可能引入估计误差。
|
||||
|
||||
## 与传统视角的对比
|
||||
|
||||
| 传统视角 | 状态估计视角 |
|
||||
|---------|------------|
|
||||
| 上下文 = 信息窗口 | 上下文 = Agent 对任务状态的信念 |
|
||||
| 目标:装更多 token | 目标:最小化信念与真实状态间的散度 |
|
||||
| 评估:token 利用率 | 评估:是否防止了下游操作错误 |
|
||||
|
||||
## 关键问题
|
||||
|
||||
- 量化每次压缩/检索/遗忘步骤中的信息损失
|
||||
- 界定 Agent 内部状态与真实任务状态的散度上界
|
||||
- 不确定性感知摘要:不仅记住内容,还记住不确定性
|
||||
- 从持久化产物重建状态而非信任自身压缩历史
|
||||
|
||||
## 与记忆评估的关联
|
||||
|
||||
记忆策略不应仅按**召回准确率**评判,还应按**是否防止了多会话任务中的下游操作错误**来评判。这需要将 [[verification-evaluation]] 与 [[context-management]] 更紧密耦合。
|
||||
|
||||
## 相关概念
|
||||
|
||||
- [[context-management]]
|
||||
- [[reliable-state-long-running-agents]]
|
||||
- [[verification-evaluation]]
|
||||
- [[agent-harness-engineering-survey]]
|
||||
Reference in New Issue
Block a user