QQWorld: Quantile-Quantile Matching for World Model Regularization
The paper argues EP regularization misses heavy-tailed latent outliers, and replaces it with rank-matched Gaussian quantiles.
QQWorld aligns projected latent samples to Gaussian quantiles so tail samples still receive corrective gradients. The authors add a cross-batch version that ranks against detached samples from earlier batches, with a stated bias-variance trade-off. In four control environments, they report higher average planning success for LeWM, better Gaussian alignment, and thinner latent tails. HF Daily Papers' note
QQWorld aligns projected latent samples to Gaussian quantiles so tail samples still receive corrective gradients. The authors add a cross-batch version that ranks against detached samples from earlier batches, with a stated bias-variance trade-off. In four control environments, they report higher average planning success for LeWM, better Gaussian alignment, and thinner latent tails. HF Daily Papers' note
score 4