Paper claims the third pretraining axis is freedom, not exploration

teortaxesTex · x · 2026-08-17

A preprint by Michael Timothy Bennett argues that the 'third axis' in generative pretraining is freedom of function space, not exploration as claimed by Explorative Modeling (XM). Experiments show that maximizing freedom maximizes generalization probability and that using freedom to guide XM training produces better, more varied images.

Original post →

More from Research

Research channel →