发表机构
Vishwakarma Institute of Technology（维斯瓦卡玛理工学院）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
研究视频预测器和世界模型中多时间尺度潜在一致性权重λ对过渡几何的作用，通过实验测量L20,q95和E20等指标，发现λ能改善移动MNIST数据集相关指标，且软一致性作用有域限制，还得出随机强迫定律统一控制域。
AI 中文摘要
多时间尺度潜在一致性是视频预测器和世界模型中常用的训练旋钮，但从业者很少知道它对过渡几何有何作用。我们将多步潜在一致性的权重λ视为诊断控制，并测量经验扩展代理L20,q95以及时间尺度为20的预测误差E20。在移动MNIST数据集（关键对处有6个种子）上，将λ从0提高到0.8可使L20从4.96±2.01降至1.01±0.06（配对t检验p = 0.005，威尔科克森检验p = 0.031），并使E20减半（从0.365降至0.177，配对t检验p = 1.1e - 13）。6个种子中有4个在λ = 0.8时L<1。相同的损失在动作条件下的摆锤-v1或推车摆杆-v1以及KTH动作视频上即使E20有所改善也不会产生总体L<1。对MMNIST的关联中介分析得出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、外部压力、世界测试、MPC、缩放）大多支持一个狭义的观点：软一致性可以将被动视频推向近收缩带，且该带受域限制。在λ = 0.8时的随机强迫定律L20 ~ 1.23 + 1.82 eta（自举斜率置信区间[1.73, 1.92]，R^2 = 0.96）通过校准的eta_eff统一了同一条曲线上的控制域。在λ为{0.4, 1.2}时的完整联合切片（30/30个单元格，5个eta x 3个种子）显示出可比的线性L20(eta)斜率（~1.69和~2.00）；我们没有拟合连续的(λ, eta)曲面。我们未报告DreamerV3或TD-MPC2的回报。
英文摘要
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.
Comments22 pages, 9 figures. Diagnostic study of multi-horizon latent consistency, expansion proxies (L20), and a stochastic-forcing law for world-model geometry. Code and experiment logs to be released publicly
