Because of its importance in academic achievement, especially in mathematics, training cognitive flexibility at school is a major issue. The present research investigates the effectiveness of a school-based intervention to improve proportion arithmetic problem solving. The study was conducted with 5th graders of 10 classes from 5 high-priority education schools in the Paris region. Students of the control and experimental groups took part in 8 learning sessions about proportion problem solving. The experimental group's training focused on comparing and flexibly categorizing the problems in the hopes to help students achieve a deeper understanding of proportion problems. Results show that training flexible categorization allowed the experimental group to progress more than the control group, in both categorization and solving tasks. The educational implications of our results are discussed.