REINFORCEMENT LEARNING TEACHERS OF TEST TIME SCALING

Summary from the TLDR AI newsletter

Using "teacher" models that explain solutions rather than solve problems from scratch, a small 7B-parameter model outperformed DeepSeek R1's 671B parameters (26.3% vs 18.9% on math benchmarks). Unlike traditional distillation where massive models must first learn to solve problems themselves, these teachers are given both questions and correct answers upfront and trained only to generate clear explanations that help student models understand.