Note COT models produce keywords like "Hmm", "Wait" during their reasoning phase. This paper supresses such words, reducing the total output token by 27-50%. Very simple approach imo, falls in "Why didn't i think of this before" category.