Slim-SC: Thought Pruning for Efficient Scaling with Self-Consistency