A class of solution methods for general large-scale systems of nonlinear equations with sparse Jacobian is proposed. In general, such a procedure uses for an update on each step not only an approximate Newton direction, but some low-dimensional subspace which may also contain other vectors such as some of the previous update vectors. The next iterate is then defined by minimization of the residual norm over the low-dimensional affine subspace. Several choices of search subspaces and strategies to perform minimization over them are proposed, analyzed, and compared numerically on a set of test problems.