QQWorld Boosts World Model Success Rate by 5.33% with Under 10 Lines of Code

In March 2026, Yann LeCun et al. released LeWorldModel (LeWM), which uses a joint‑embedding predictive architecture (JEPA) and Gaussian regularization to learn environment dynamics from raw pixels. Despite Gaussian regularization, the latent variable distribution still exhibits heavy‑tail behavior, where a few latent variables drift far from the main distribution.

The authors identify that the Epps–Pulley (EP) normality objective employed by LeWM suffers from a "tail blind spot": it can detect overall deviation from a Gaussian but provides diminishing corrective gradients for extreme latent samples, making it hard to pull them back into the core latent space.

To address this, the XJTU research team proposes QQWorld, which replaces the EP regularizer with quantile‑quantile (QQ) matching. Instead of indirect kernel‑based distance, QQWorld sorts the random projections of latent variables and directly moves each sample toward the corresponding quantile of a standard Gaussian, ensuring persistent gradient signals even for tail samples.

The method computes the order statistics of projected samples, maps each rank to the matching Gaussian quantile, and minimizes the squared distance between them. This per‑sample target yields stronger corrective forces for outliers, effectively suppressing heavy‑tail latent variables.

The paper also proves a one‑way control relationship: when the QQ loss approaches zero, the EP discrepancy also vanishes, but the converse does not hold, highlighting QQ regularization’s superiority.

Because QQ matching relies on batch‑wise ordering, the authors introduce Cross‑Batch QQ, which maintains a FIFO queue of recent projection features to enlarge the ordering pool without increasing back‑propagation memory. Theoretical analysis shows a bias‑variance trade‑off: moderate queue lengths reduce variance of quantile estimates, while overly long queues introduce stale‑distribution bias.

Experiments on Four control environments (Two‑Room, Reacher, PushT, OGBench‑Cube) show that QQWorld raises average planning success from 79.75% to 85.08% (+5.33 points). Latent distribution metrics improve: QQ RMSE drops from 0.157 to 0.121, tail probability from 0.315 to 0.123, and trajectory reconstruction RMSE from 6.01 to 2.58.

Memory ablation demonstrates that using a physical batch size of N=32 with two historical batches achieves 83.50% success while using only ~34% of the GPU memory required by the baseline QQWorld with N=128, incurring only a 1.58‑point success loss.

In summary, QQWorld replaces EP regularization with direct quantile matching, providing effective gradient correction for extreme latent samples, improving both planning performance and latent space regularity. Cross‑Batch QQ further decouples batch size from memory consumption, enabling scalable training of world models.