The focus of this investigation is on the renewed assessment of nonlinear reduced order models (ROM) for the accurate prediction of the geometrically nonlinear response of a curved beam. In light of difficulties encountered in an earlier modeling effort, the various steps involved in the construction of the reduced order model are carefully reassessed. The selection of the basis functions is first addressed by comparison with the results of proper orthogonal decomposition (POD) analysis. The normal basis functions suggested earlier, i.e. the transverse linear modes of the corresponding flat beam, are shown in fact to be very close to the POD eigenvectors of the normal displacements and thus retained in the present effort. A strong connection is similarly established between the POD eigenvectors of the tangential displacements and the dual modes which are accordingly selected to complement the normal basis functions. The identification of the parameters of the reduced order model is revisited next and it is observed that the standard approach for their identification does not capture well the occurrence of snap-throughs. On this basis, a revised approach is proposed which is assessed first on the static, symmetric response of the beam to a uniform load. A very good to excellent matching between full finite element and ROM predicted responses validates the new identification procedure and motivates its application to the dynamic response of the beam which exhibits both symmetric and antisymmetric motions. While not quite as accurate as in the static case, the reduced order model predictions match well their full Nastran counterparts and support the reduced order model development strategy.