Tag
Introduces PROPEL, a solver-amortized framework that trains a lightweight activation probe to predict solver pass rates, enabling efficient training of task generators for RL without costly solver rollouts. The method improves generation at the learnable frontier across math, code, and software-engineering tasks.