20260625:很多新内容
This commit is contained in:
44
concepts/rollout-drift.md
Normal file
44
concepts/rollout-drift.md
Normal file
@@ -0,0 +1,44 @@
|
||||
---
|
||||
title: "Rollout Drift (推演漂移)"
|
||||
created: 2026-06-23
|
||||
updated: 2026-06-23
|
||||
type: concept
|
||||
tags: ["world-modeling", "latent-prediction", "error-accumulation", "regularization"]
|
||||
sources: ["[[hazare-dcgwm-2026]]", "https://arxiv.org/abs/2606.18688", "GIRL (2026)"]
|
||||
---
|
||||
|
||||
# Rollout Drift (推演漂移)
|
||||
|
||||
**Rollout drift** 是多步潜在预测中的累积误差现象:在自回归展开过程中,每步预测的微小误差逐步叠加,使预测轨迹偏离接地流形。
|
||||
|
||||
## 问题的本质
|
||||
|
||||
[[world-models-rl|世界模型]]中的潜在状态 z_t 通过预测函数 f 向前展开:
|
||||
|
||||
```
|
||||
z_{t+1} = f(z_t, a_t) + ε_t
|
||||
```
|
||||
|
||||
每步预测误差 ε_t 在后续步骤中被放大——因为下一步的输入 z_{t+1} 包含了误差,而 f 是非线性的。在长视距 (T ≫ 1) 下:
|
||||
|
||||
```
|
||||
z_T = f(f(...f(z_0, a_0)..., a_{T-2}), a_{T-1}) + 累积误差
|
||||
```
|
||||
|
||||
## 为什么特别棘手
|
||||
|
||||
GIRL (2026) 将 rollout drift 识别为潜在世界模型在长视距下的**中心失效模式**。标准方案(如单步对齐损失 L_PGC、L_SBGC)仅在训练时的单步上接地,不能保证多步展开轨迹保持在接地上。
|
||||
|
||||
## DCGWM 的解决方案:L_AGA
|
||||
|
||||
[[asymmetric-grounding-adherence-loss|Asymmetric Grounding Adherence Loss (L_AGA)]] 是首个针对异质接地源、具有不兼容容忍结构的 rollout drift 损失:
|
||||
|
||||
- **物理漂移** → 平方铰链惩罚:违反物理约束是范畴错误,超过容忍阈值后二次惩罚
|
||||
- **行为漂移** → 软 KL 散度:行为随机性是预期内的,使用与分布距离成比例的连续惩罚
|
||||
|
||||
## 参考
|
||||
|
||||
- [[hazare-dcgwm-2026|DCGWM 论文]]
|
||||
- [[asymmetric-grounding-adherence-loss|L_AGA]]
|
||||
- [[world-models-rl|World Models]]
|
||||
- [[objective-interference-collapse|OIC]]
|
||||
Reference in New Issue
Block a user