Files
myWiki/concepts/contextual-transformation-field.md
2026-07-20 14:14:55 +08:00

63 lines
1.9 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: "Contextual Transformation Field (上下文变换场)"
created: 2026-07-10
updated: 2026-07-10
type: concept
tags: ["representation-learning", "vector-field", "contextual-representation", "llm-geometry"]
sources: ["[[shared-concept-geometry-2026|Hu et al. (ICML 2026)]]"]
---
# Contextual Transformation Field (上下文变换场)
Hu et al. (ICML 2026) 的核心创新——将 LLM 中概念在上下文间的位移形式化为**向量场**。
## 定义
设 τ₀ 为源上下文(如中性模板),τ 为目标上下文。概念 w 的上下文位移:
```
φ^{(m)}(w, τ) = r^{(m)}(w, τ) - r^{(m)}(w, τ₀) ∈ M^{(m)}
```
跨词汇 W 的位移集合构成向量场:
```
Φ_τ^{(m)} : W → M^{(m)}, w ↦ φ^{(m)}(w, τ)
```
由其核矩阵表征:
```
K_Φτ [i, j] = k(φ^{(m)}(w_i, τ), φ^{(m)}(w_j, τ))
```
## 关键发现
### 场不是均匀的
- **主导方差比** ρ₁ ∈ [0.073, 0.146]——远未达到单一方向主导
- **球面全局方差** V_global ∈ [0.43, 0.75] rad²均匀分布 ~2.47)——场分散但非随机
### 方差是语义组织的
| 词属性 | 与 | 相关性 | p 值 |
|--------|-----|--------|------|
| 词汇密度 | 位移幅度 r̃ | ρ ∈ [-0.22, -0.32] | < 0.001 |
| 具体性 | 方向偏差 σ | ρ [-0.21, -0.41] | < 0.001 |
语义邻域密集的词位移更小更具体的概念方向偏差更小抽象概念因缺乏感知锚定在语义框架间被推向更分散的方向
### 跨模型共享
位移关系结构在模型间可迁移从模型 A 运送位移结构到模型 B预测保留位移显著高于基线
## 对表示工程的意义
单向量 steering Arditi et al., 2024将上下文变换场近似为均匀平移 **丢弃了语义上有意义的结构化残差**
## 参考
- [[shared-concept-geometry-2026|Hu et al. (ICML 2026)]]
- [[concept-point-cloud-manifold|Point-Cloud Manifold]]
- [[within-concept-between-concept-axes|Two Axes]]