Implicit Curriculum in RL with Verifiable Rewards

A new theory explains how reinforcement learning with verifiable rewards (RLVR) enables transformers to solve compositional reasoning tasks. The theory reveals that mixed-difficulty training creates an implicit curriculum: easier problems become learnable first, shaping the frontier for harder ones. This learning progression from easy to hard emerges without any explicit schedule. The effectiveness depends on the smoothness of the difficulty spectrum. When smooth, a relay regime occurs where gradient signals on easier problems make slightly harder ones tractable. The research is published on arXiv (2602.14872v2).