Xi'an Jiaotong's QQWorld Boosts World Model Planning Success Rate

机器之心 · wechat · 2026-08-11

To address the heavy-tail issue in the latent space of Yann LeCun's LeWorldModel (LeWM), researchers from Xi'an Jiaotong University proposed QQWorld. They found that the EP normality objective used in LeWM has a 'tail blind spot,' struggling to correct extreme latent variables that drift far from the main distribution.

QQWorld replaces the EP regularizer with Quantile-Quantile Matching, directly moving each projected sample to its corresponding standard Gaussian quantile. This provides continuous effective gradients at the distribution tails, with a core implementation of less than 10 lines of code.

Experiments show QQWorld increases the average planning success rate from 79.75% to 85.08% across multiple control environments. Additionally, the proposed Cross-Batch QQ mechanism uses historical batch features to stabilize ordering, achieving significantly better performance than the baseline while reducing physical batch size to a quarter and cutting GPU memory by 73%.

Original post →

More from Research

Research channel →