We propose a compositional energy-based approach with a new sampling method (PEM) that improves generalization on complex reasoning tasks.