From what I understood, these models differ from CARTs for regression, mostly because they fit a linear model at the leaves of the tree instead of simply taking an average. They also "smooth" the tree by generating linear models in the intermediate steps of the tree growth process.
I have been using the R implementation of them a bit for regression getting very good results. But I wonder about the assumptions of a usual linear model? Multicolinearity, Autocorrelation, Non-Normality being violated doesnt worry people in this case?