Files
myWiki/concepts/insight-backpropagation.md

38 lines
1.3 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: "Insight Backpropagation"
created: 2026-06-24
updated: 2026-06-24
type: concept
tags: ["knowledge-management", "agent", "tree-structure", "abstraction"]
sources:
- "[[arbor-htr-2026]]"
---
# Insight Backpropagation
Insight Backpropagation 是 Arbor 假设树中从叶到根的洞察传播机制:将局部实验结果抽象为方向级经验,最终贡献到全局紧凑理解。
## 工作流
1. **叶子执行**Executor 返回 (score, result, insight, branch)
2. **写回节点**:将结果写入对应叶子节点
3. **向上抽象**:沿祖先路径,每个内部节点 Abstract({ι_c for c in children})
4. **全局汇聚**:根的 insight 汇总整次运行的 compact understanding
## 洞察 vs 日志
| 维度 | 执行日志 | Insight |
|------|---------|---------|
| 内容 | 原始工具调用/输出 | 语义抽象 |
| 示例 | "Ran Muon with lr=0.01, loss=0.043" | "轴统计量单独不够;替换 NS 会破坏全矩阵几何" |
| 复用性 | 低(需人工解读) | 高(直接约束后续假设生成) |
## 为什么需要
没有抽象,树会退化为无结构日志——后续的 Ideate 和 Select 无法从历史中学习。Insight Backpropagation 是 HTR 从"探索日志"升级为"累积研究"的关键。
## 参考
- [[arbor-htr-2026]]
- [[hypothesis-tree-refinement]]
- [[coordinator-executor-architecture]]