Actually, Matlab does not use Arpack anymore. Instead, they have implemented their own version of the Krylov-Schur method accessible from the Matlab prompt with edit eigs.m.
I am not sure whether the following tweak will resolve your issue, however, I remember from my own work on a modesolver that SuiteSparse conducts an iterative refinement of the solutions by default. Disabling this refinement using
SuiteSparse.UMFPACK.umf_ctrl[8] = 0
improved the performance of eigs significantly for my use case.