Xi'an Jiaotong University Proposes QQWorld: Just 10 Lines of Code, World Model Success Rate Improves by 5.33 Percentage Points

Researchers at Xi'an Jiaotong University propose QQWorld, a method that replaces the original EP regularization with quantile-quantile matching, effectively suppressing heavy tails in the latent variable distribution of world models. This improves the average planning success rate from 79.75% to 85.08%, with core code under 10 lines. The method also introduces Cross-Batch QQ, which reduces memory usage by about 73% while maintaining performance, offering an efficient solution for large-scale world model training.