In the application of an interior point method to certain convex smooth nonlinear programming problems numerical difficulties in the neighborhood of the solution had to be observed. The reason are cancellation errors in the computation of active components. These errors are not critical for the computation of the variables themselves, but they are also used for some right hand sides in subsequent computations and then lead to problems in stable computation of a solution. There are known techniques to eliminate such critical variables. In this paper, a new possibility to remove this problem is proposed, with the advantage that knowledge on the corresponding dual variables is not lost. It is outlined that an underlying implementation can be changed easily to handle the problem above.