we consider methods for large-scale unconstrained optimization based on finding an approximate solution of a quadratically constrained trust-region subproblem. the solver is based on sequential subspace minimization with a modified barrier "accelerator" direction in the subspace basis.