SAGE introduces a self-aware sampling approach allowing large reasoning models to decide when to stop thinking, reducing unnecessary computation while improving accuracy.