QQWorld: Fixing Heavy-Tailed Deviations in World Models with Quantile-Quantile Matching

Zhoushun Yu · hf · 2026-08-03

To address the issue of vanishing corrective gradients for tail samples in latent world models using the EP objective, researchers propose QQWorld.

This method replaces EP with a quantile-quantile matching objective that directly aligns projected latent samples with rank-matched Gaussian quantiles, maintaining effective corrective gradients in the tails. They also develop cross-batch QQ to enlarge the effective ranking pool using detached samples from previous batches, characterizing its bias-variance trade-off.

Experiments across four control environments show that QQWorld effectively improves the average planning success rate of LeWM, consistently yielding better Gaussian alignment and thinner latent tails.

Original post →

More from Research

Research channel →