Combining multiple tokens into fewer tokens to reduce computation while preserving model output quality.