One of the advantages, as stated in the paper you linked, is that with QAOA you can increase the precision arbitrarily, whereas QA will only find the solution with probability 1 as $T \to \infty$ which is impractical. In addition if $T$ is too long you're likely to not find the solution as the probability is not monotonic. I believe an example of this can be found in a fair-sampling paper by Matsuda et al. Figure 4 shows that for large $\tau$, using quantum annealing on a 5-qubit system, you only find 2 of the 3 possible states. 

Matsuda et al. https://arxiv.org/abs/0808.0365v3