In a set of nonlinear models, we propose reparametrization that facilitates robust inference under weak identification.