Freedom Selection Outperforms MDL by up to 500% in Induction
August 7, 2026
Analyzing Explorative Modeling (XM), this work reframes the third pretraining axis as behavioral freedom rather than mere expressivity. Formal proofs and induction experiments show that selecting for the weakest models maximizes generalization, with freedom selection beating Minimum Description Length (MDL) by 110–500%.
HOW THIS AFFECTS YOU
●
researcherYou can potentially improve generalization by optimizing for behavioral freedom rather than just MDL or expressivity.