44 lines
1.4 KiB
Markdown
44 lines
1.4 KiB
Markdown
---
|
||
title: "Hypothesis Tree Refinement (HTR)"
|
||
created: 2026-06-24
|
||
updated: 2026-06-24
|
||
type: concept
|
||
tags: ["autonomous-research", "agent", "tree-search", "knowledge-management"]
|
||
sources:
|
||
- "[[arbor-htr-2026]]"
|
||
---
|
||
|
||
# Hypothesis Tree Refinement (HTR)
|
||
|
||
HTR 是 Arbor 框架的核心方法：将自主科研的中间状态持久化为假设树，通过分支探索、执行反馈回写、洞察向上传播和 held-out 准入门控实现累积式科研进步。
|
||
|
||
## 树节点结构
|
||
|
||
n = ⟨h, ι, µ⟩
|
||
- **h (Hypothesis)**：可验证的改进主张，粒度随深度细化（根=方向，叶=具体干预）
|
||
- **ι (Insight)**：可复用的证据解读——非执行日志，紧致语义记忆
|
||
- **µ (Metadata)**：状态/分数/git ref
|
||
|
||
## 五步循环
|
||
|
||
```
|
||
Observe → Ideate → Select → Dispatch → Backpropagate
|
||
```
|
||
|
||
1. **Observe**：观察当前树状态（前沿、洞察、约束）
|
||
2. **Ideate**：在选定父节点下生成 k 个子假设
|
||
3. **Select**：选择最有前景的叶子调度执行
|
||
4. **Dispatch**：将叶子分配给隔离 Executor
|
||
5. **Backpropagate**：将执行结果（分数/洞察）写回节点，沿祖先路径向上抽象
|
||
|
||
## 三种角色合一
|
||
|
||
- **搜索前沿**：活跃/验证/剪枝方向的可视化
|
||
- **长期记忆**：成功+失败的复用证据
|
||
- **可审计记录**：每个产物变更可追溯到动机假设
|
||
|
||
## 参考
|
||
- [[arbor-htr-2026]]
|
||
- [[coordinator-executor-architecture]]
|
||
- [[autonomous-optimization-ao]]
|