It removes $S_b^2$ from the error term of the treatment comparison.

It increases the degrees of freedom for the treatment effect.

It eliminates the need for randomization within blocks.

It ensures that $S_w^2$ is zero.