In this paper we enhance the adaptive scheme presented in Cavoretto and De Rossi (2019) for solving elliptic boundary value problems via RBF collocation methods. More precisely, this study concerns a leave one out cross validation technique applied as an error estimate and used in the adaptive refinement process. The modified algorithm we propose here allows us to get numerical convergence also when L-shape or irregular domains are considered. Moreover, a comparison between unsymmetric and symmetric RBF collocation schemes is performed.