This paper presents a hybrid pronoun reso- lution system for German. It uses a simple rule-driven entity-mention formalism to in- crementally process discourse entities. An- tecedent selection is performed based on Markov Logic Networks (MLNs). The hy- brid architecture yields a cheap problem formulation in the MLNs w.r.t. inference complexity but pertains their expressive- ness. We compare the system to a rule- driven baseline and an extension which uses a memory-based learner. We find that the MLN hybrid outperforms its competi- tors by large margins.