“Experiments with s1 models showed that simple rejection sampling to control Chain-of-Thought length leads to reversed scaling, where longer CoTs result in worse performance.”