A new way of constructing efficient semiparametric instrumental variable estimators is proposed. The method involves the combination of a large number of possibly inefficient estimators rather than combining the instruments into an optimal instrument function. The consistency and asymptotic normality is established for a class of estimators that are linear combinations of a set ofv?? ?? consistent estimators whose cardinality increases with sample size. It is shown that the semiparametrically efficient estimator lies in this class. The proofs do not rely on smoothness of underlying criterion functions. Potential use of the estimator can overcome the undersized sample problem. in simultaneous equation system estimation.