Mostly because it's impractical, sometimes because they are lazy or it's simply not statistically significant.
If you train a very large NN it's often to expensive to do it several times. And on very large validation sets you really get significant results pretty fast, so there is not really a need for it. However, I agree that some minor permutations of the NN architecture is often just noise and groups publish it for the sake of publishing.
Mostly because it's impractical, sometimes because they are lazy or it's simply not statistically significant.
If you train a very large NN it's often to expensive to do it several times. And on very large validation sets you really get significant results pretty fast, so there is not really a need for it. However, I agree that some minor permutations of the NN architecture is often just noise and groups publish it for the sake of publishing.