Higher-than-second-order statistics-based input/output identification algorithms are proposed for linear and nonlinear system identification. The higher-than-second-order cumulant-based linear identification algorithm is shown to be insensitive to contamination of the input data by a general class of noise including additive Gaussian noise of unknown covariance, unlike its second-order counterpart. The nonlinear identification is at least as optimal as any linear identification scheme. Recursive-least-squares-type algorithms are derived for linear/nonlinear adaptive identification. As applications, the problems of adaptive noise cancellation and time-delay estimation are discussed and simulated. Consistency of the adaptive estimator is shown. Simulations are performed and compared with the second-order design.