QQWorld Replaces Epps-Pulley Regularizer with Quantile-Quantile Matching to Fix Tail-Latent Control in World ModelsQQWorld：用分位数-分位数匹配替代Epps-Pulley正则项，修复世界模型隐空间尾部失控问题
S 1.7T11 sources1 个来源R7-research
Researchers show that LeWorldModel (LeWM), which regularizes its latent representations toward an isotropic Gaussian using the Epps-Pulley (EP) objective for efficient world-model planning, suffers because EP's corrective gradients rapidly vanish for isolated tail samples, leaving heavy-tailed latent deviations insufficiently controlled.
To address this, Zhoushun Yu, Xiaoyu Hu, and Xiangyu Xu (arXiv:2607.28415v1, submitted July 30, 2026) propose QQWorld, which replaces the EP objective with a quantile-quantile matching objective that directly aligns projected latent samples to the target Gaussian's quantiles, giving more direct control over tail behavior.
Latent world models underpin planning in autonomous systems, so more reliably regularized latent spaces could improve the robustness of trajectory and behavior prediction used in autonomous-driving planning stacks, particularly for rare, tail-distribution driving scenarios.