ACT as a Conditional VAE encoder

Thanks for the research and open-source. 
I am wondering regarding: the C-VAE which seems to be beneficial when there is one task being learned but is it still useful and valid when multi-task is introduced? especially that the task embedding doesn't go into the CVAE encoder? there supposed to be a difference between task distribution or so-called style (z).