QQWorld Improves Latent World Model Regularization via Quantile Matching
July 31, 2026
QQWorld replaces the Epps-Pulley objective with quantile-quantile matching to prevent vanishing gradients in the tails of latent distributions. By aligning projected samples with rank-matched Gaussian quantiles, it maintains effective regularization for latent world models in control environments.
HOW THIS AFFECTS YOU
●
researcherThis method provides more stable latent distribution modeling for world models by addressing heavy-tailed deviations.