World models depend on well-behaved latent distributions: heavy tails or isolated outliers in latent space can break downstream planning even when global statistics look normal. QQWorld's core insight is that the commonly used Epps–Pulley (EP) objective produces vanishing corrective gradients for isolated tail samples, so those deviations remain uncontrolled; replacing EP with a quantile–quantile (QQ) matching objective preserves effective corrective gradients in the tails and yields thinner latent tails and better Gaussian alignment.
Key Findings
- QQ matching aligns one-dimensional projections of latent samples to rank-matched Gaussian quantiles, which maintains non-vanishing corrective gradients for tail samples — so outliers are actively corrected rather than ignored.
- Cross-batch QQ enlarges the effective ranking pool by using detached samples from previous batches, reducing rank noise and stabilizing the objective at the cost of a bias–variance trade-off that can be tuned.
- Empirically replacing EP with QQ in LeWorldModel improves average planning success across four control environments and consistently yields tighter Gaussian alignment and thinner latent tails — so planning and control become more reliable under the same latent capacity.
Who it's for and tradeoffs
Great fit if you develop latent world models or planning agents and face brittle downstream planning due to heavy-tailed latents. QQWorld is especially relevant when per-batch ranking is noisy or datasets are small, since cross-batch ranking can increase effective sample size. Look elsewhere or be cautious if your latents are already well-Gaussianized, when additional ranking bookkeeping is undesirable, or when strict unbiasedness of the regularizer is required — cross-batch detachment introduces a tunable bias–variance trade-off.
Method sketch
Project latent vectors onto scalar directions, sort projected samples to obtain empirical ranks, and match each rank to the corresponding Gaussian quantile; optimize a QQ loss that penalizes displacement from these quantiles. Cross-batch QQ augments ranking using detached stored samples from prior batches to reduce rank variance while preventing gradient leakage into past states. This simple change retains anti-collapse behavior without relying on EP's kernel-based statistics, and keeps gradients meaningful in the tails.