improve thinking-output for mulit-variant