Files
myWiki/concepts/trajectory-regulation-layer.md

46 lines
1.7 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: "Trajectory Regulation Layer轨迹调控层"
created: 2026-06-11
updated: 2026-06-11
type: concept
tags: [agent, harness, trajectory, recovery, monitoring]
sources: [raw/papers/xu-life-harness-runtime-adaptation-2026.md]
---
# Trajectory Regulation Layer
> Life-Harness 的第④层:环境反馈返回后,监控轨迹级模式,检测**退化信号**并在必要时触发恢复干预。
## 机制
```
r_t = REGULATE_TRAJECTORY(τ_t, a_t, o_{t+1}, b_t)
```
其中 b_t = B - t - 1 是剩余步骤预算。输出 r_t 可能是:
- **空** — 轨迹正常,无需干预
- **软恢复消息** — 轻微偏离,提示纠正
- **重复失败警告** — 检测到模式重复
- **强纠正指令** — 轨迹明显退化,需要强制重定向
## 检测的退化模式
该层不依赖于深度语义理解,而是从**轨迹级模式**中检测问题:
- **重复Repetition**:连续发出相同的无效命令
- **停滞Stagnation**:在等价状态间循环,无实质进展
- **无效重试Invalid Retries**:反复重试已知失败的操作
- **预算耗尽Budget Exhaustion**:剩余步骤不足以完成任务
## 设计哲学
这类失败是**自强化self-reinforcing**的——Agent 陷入循环后不会自行跳出。它们通常不需要深度推理即可从轨迹模式中检测,但需要**外部介入**才能打断。这正是 Regulation Layer 的价值。
在 Life-Harness 的失败诊断中,轨迹退化占 17%。
## 参考
- [[xu-life-harness|Life-Harness 论文]]
- [[action-realization-layer|动作实现层]]
- [[deterministic-agent-failures|确定性 Agent 失败分类]]
- [[harness-evolution|Harness Evolution]]