20260617:目前有914 页
This commit is contained in:
44
concepts/runtime-harness-adaptation.md
Normal file
44
concepts/runtime-harness-adaptation.md
Normal file
@@ -0,0 +1,44 @@
|
||||
---
|
||||
title: "Runtime Harness Adaptation(运行时骨架适配)"
|
||||
created: 2026-06-11
|
||||
updated: 2026-06-11
|
||||
type: concept
|
||||
tags: [agent, harness, runtime-adaptation]
|
||||
sources: [raw/papers/xu-life-harness-runtime-adaptation-2026.md]
|
||||
confidence: high
|
||||
---
|
||||
|
||||
# Runtime Harness Adaptation
|
||||
|
||||
> 不更新模型权重,而是进化模型与确定性环境之间的 Runtime 接口层——将训练轨迹中的反复失败转化为可复用的结构化干预。
|
||||
|
||||
## 核心定义
|
||||
|
||||
Runtime Harness Adaptation 是一种 **Agent 改进范式**:保持 LLM 参数冻结(θ fixed),通过适配运行时骨架 H′ ← A_harness(H, T_train) 来提升 Agent 的交互表现。适配后的 harness 在评估时保持固定,不修改模型权重,也不修改评估环境。
|
||||
|
||||
与 [[runtime-interface-adaptation|参数适配(Parameter Adaptation)]] 的区别在于:参数适配将领域特定的行为吸收进模型权重,需要随模型或环境变化重新训练;Harness 适配捕获的是**环境侧的可复用结构**,具有模型无关性。
|
||||
|
||||
## 四个生命周期层
|
||||
|
||||
Life-Harness 将 Harness 适配组织为四个生命周期阶段(详见 [[xu-life-harness|Life-Harness 论文]]):
|
||||
|
||||
1. **环境契约层** [[environment-contract-layer]] — 交互前校准工具和接口约束
|
||||
2. **程序技能层** [[procedural-skill-layer]] — 从轨迹中检索可复用程序
|
||||
3. **动作实现层** [[action-realization-layer]] — 执行前验证和规范化动作
|
||||
4. **轨迹调控层** [[trajectory-regulation-layer]] — 执行后检测退化并触发恢复
|
||||
|
||||
## 为什么有效
|
||||
|
||||
在确定性、规则驱动的领域中,大量 Agent 失败不是模型"不够聪明",而是接口层的不匹配:
|
||||
- 工具契约被误解
|
||||
- 动作格式不可执行
|
||||
- 反馈无法触发恢复
|
||||
- 轨迹陷入重复/停滞
|
||||
|
||||
这些失败是**可诊断、可结构化、可跨模型复用的**——这正是 Harness 适配的用武之地。
|
||||
|
||||
## 参考
|
||||
- [[xu-life-harness|Life-Harness 论文]]
|
||||
- [[agent-harness-engineering|Agent Harness Engineering]]
|
||||
- [[agent-harness-mini|Mini Agent Harness]]
|
||||
- [[cross-model-harness-transfer|跨模型 Harness 迁移]]
|
||||
Reference in New Issue
Block a user