Consider a semiparametric regression modelY=f(θ, X, ϵ), wherefis a known function,θis an unknown vector, ϵ consists of a random error and possibly of some unobserved variables, and the distributionF(·) of (ϵ,X) is unspecified. This article introduces, in a general setting, new methodology for estimatingθandF(·). The proposed method constructs a profile likelihood defined on random-level sets (a random sieve). The proposed method is related to empirical likelihood but is more generally applicable. Four examples are discussed, including a quadratic model, high-dimensional semiparametric regression, a nonparametric random-effects model, and linear regression with right-censored data. Simulation results and asymptotic analysis support the utility and effectiveness of the proposed method.