When using glmmLasso.predict with a model that was fitted using family=acat() and more than two categories, predictions are wrong because the dependent variable does not get transformed to create multiple intercepts. This probably generalizes to all types of multivariate models but has only been tested with family=acat().
This might be a bug but needs to be confirmed with a minimal working example. @gdepalo came across this issue, so she should be able to provide the MWE.