20260625:很多新内容
This commit is contained in:
63
concepts/dcgwm.md
Normal file
63
concepts/dcgwm.md
Normal file
@@ -0,0 +1,63 @@
|
||||
---
|
||||
title: "DCGWM (双通道接地世界建模)"
|
||||
created: 2026-06-23
|
||||
updated: 2026-06-23
|
||||
type: concept
|
||||
tags: ["world-modeling", "jepa", "representation-learning", "architecture"]
|
||||
sources: ["[[hazare-dcgwm-2026]]", "https://arxiv.org/abs/2606.18688"]
|
||||
---
|
||||
|
||||
# DCGWM (Dual-Channel Grounded World Modeling)
|
||||
|
||||
**DCGWM** 是 Hazare (2026) 提出的世界模型架构,通过**分区潜在空间 + 内向梯度流**结构性防止 [[objective-interference-collapse|Objective Interference Collapse]]。
|
||||
|
||||
## 架构核心
|
||||
|
||||
```
|
||||
物理测量 → PGC [内向∇] → Z_p ┐
|
||||
├→ LWME → detach() → GRL → 用户
|
||||
社交模拟 → SBGC [内向∇] → Z_b ┘
|
||||
```
|
||||
|
||||
### 组件
|
||||
|
||||
1. **Latent World Modeling Engine (LWME)**:基于 [[jepa|JEPA]] 的预测引擎,潜在空间 Z = Z_p ⊕ Z_b 通过架构强制分区(独立权重组、独立 LayerNorm、互信息最小化)
|
||||
2. **Physical Grounding Channel (PGC)**:通过 [[vicreg|VICReg]] 风格对齐将外部物理测量接地到 Z_p,[[inward-only-gradient-flow|内向梯度流]] 仅更新 W_p
|
||||
3. **Social-Behavioral Grounding Channel (SBGC)**:将紧急性多智能体模拟输出作为外部接地信号注入 Z_b,内向梯度流仅更新 W_b
|
||||
4. **Inter-Channel Interface Module**:通过一致性 + 解耦的双目标耦合两子空间,无跨子空间梯度
|
||||
5. **Generative Rendering Layer (GRL)**:架构隔离的生成层,接收 detach() 的潜在表示
|
||||
|
||||
### 四个架构不变量
|
||||
|
||||
| # | 不变性 | 含义 |
|
||||
|---|--------|------|
|
||||
| I1 | Z_p, Z_b 无共享参数 | 无直接梯度路径 |
|
||||
| I2 | PGC→W_p only; SBGC→W_b only | 内向梯度流 |
|
||||
| I3 | 接口无跨子空间梯度 | 耦合但不污染 |
|
||||
| I4 | GRL detach() | 生成损失不回流 |
|
||||
|
||||
## 训练阶段
|
||||
|
||||
1. LWME 预训练(仅 L_pred)
|
||||
2. PGC 对齐(仅 W_p)
|
||||
3. SBGC 对齐(仅 W_b)
|
||||
4. 联合微调(全部项 + [[asymmetric-grounding-adherence-loss|L_AGA]])
|
||||
5. GRL 训练(LWME 冻结 + detach)
|
||||
|
||||
## 与相关工作的区别
|
||||
|
||||
- **vs. Domain Expansion**:DCGWM 的子空间锚定于**外部**接地源(而非内部任务目标),内向约束完全阻止跨子空间梯度(而非仅正交投影)
|
||||
- **vs. ThinkJEPA**:ThinkJEPA 的 VLM thinker 信号注入共享 JEPA 预测器→梯度可在两路径间传播。DCGWM 的分区潜在空间 + 内向梯度流结构性防止此干涉
|
||||
- **vs. WMReward/GIRL**:DCGWM 将物理接地作为特例,增加行为接地通道、OIC 形式化、L_AGA、隔离必要性定理
|
||||
|
||||
## 局限性
|
||||
|
||||
OIC 是猜想(非形式证明),行为编码器保真度未验证,接口收敛未证明,隔离必要性依赖未证明假设 A2,零实验验证。
|
||||
|
||||
## 参考
|
||||
|
||||
- [[hazare-dcgwm-2026|DCGWM 论文]]
|
||||
- [[objective-interference-collapse|OIC]]
|
||||
- [[inward-only-gradient-flow|Inward-Only Gradient Flow]]
|
||||
- [[asymmetric-grounding-adherence-loss|L_AGA]]
|
||||
- [[isolation-necessity-theorem|Isolation Necessity]]
|
||||
Reference in New Issue
Block a user