调节视频模型的多步一致性，可让潜变量动态趋于收缩
Multi-Horizon Consistency as Geometry: When Latent Dynamics Contract, and When They Do Not
- 用一致性权重λ控制潜变量动态，使其更趋近于收缩状态
- 在Moving-MNIST上，λ=0.8时潜变量扩张度降至1.01，预测误差减半
- 该效果仅在特定数据集有效，说明收缩能力具有领域限制
多步潜变量一致性是视频预测与世界模型中的常见训练参数，但其对转移几何的影响常被忽视。本文将一致性权重λ作为诊断工具，测量经验性扩张代理量L20,q95与20步预测误差E20。在Moving-MNIST（n=6种子，临界组合）中，λ从0增至0.8，使L20由4.96±2.01降至1.01±0.06（配对t检验p=0.005，Wilcoxon p=0.031），E20由0.365降至0.177（配对t检验p=1.1e-13）。六组中有四组在λ=0.8时实现L<1。该损失在动作条件下的Pendulum-v1、CartPole-v1及KTH Actions视频上无法使总体L<1，即便E20改善亦然。在Moving-MNIST上进行关联中介分析得r-hat=0.94（95%置信区间[0.88, 1.00]，n=27，B=2000）；λ未随机化。防御性检查（架构基线、外生压力、WorldTest、MPC、扩展性）支持有限结论：软一致性可使被动视频接近近收缩区域，且该区域受领域限制。在λ=0.8时，随机扰动律满足L20 ~ 1.23 + 1.82 eta（bootstrap斜率置信区间[1.73, 1.92]，R²=0.96），通过校准η_eff统一不同控制域于同一曲线。λ取值{0.4, 1.2}的完整联合切片（30/30单元，5个η×3个种子）显示相似线性斜率（约1.69和约2.00），未拟合连续(λ, η)曲面。未报告DreamerV3或TD-MPC2的回报。
原文摘要 · Abstract (English)
Multi-horizon latent consistency is a common training knob in video predictors and world models, but practitioners rarely know what it does to transition geometry. We treat lambda, the weight on multi-step latent agreement, as a diagnostic control and measure an empirical expansion proxy L20,q95 together with horizon-20 prediction error E20. On Moving-MNIST (n=6 seeds at the critical pair), raising lambda from 0 to 0.8 cuts L20 from 4.96 +/- 2.01 to 1.01 +/- 0.06 (paired t p=0.005, Wilcoxon p=0.031) and halves E20 (0.365 to 0.177, paired t p=1.1e-13). Four of six seeds cross L<1 at lambda=0.8. The same loss does not produce population L<1 on action-conditioned Pendulum-v1 or CartPole-v1, nor on KTH Actions video, even when E20 improves. An associational mediation analysis on MMNIST gives r-hat=0.94 (95% CI [0.88, 1.00], n=27, B=2000); lambda was not randomized. Defensive checks (architectural baselines, exogenous stress, WorldTest, MPC, scaling) mostly support a narrow claim: soft consistency can push passive video toward a near-contractive band, and that band is domain-limited. A stochastic-forcing law L20 ~ 1.23 + 1.82 eta at lambda=0.8 (bootstrap slope CI [1.73, 1.92], R^2=0.96) unifies control domains on the same curve via calibrated eta_eff. Complete joint slices at lambda in {0.4, 1.2} (30/30 cells, 5 eta x 3 seeds) show comparable linear L20(eta) slopes (~1.69 and ~2.00); we do not fit a continuous (lambda, eta) surface. We do not report DreamerV3 or TD-MPC2 returns.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。