Discovering simple, understandable and yet accurate process models is a well-known issue for models mined from real-life event logs. In this paper, we consider algorithms for automatically computing nested groups of activities in declarative process languages, concretely Dynamic Condition Response (DCR) Graphs, to reduce complexity while preserving accuracy. The DCR Graphs notation is, on the one hand, supported by the very accurate DisCoveR process mining algorithm, and on the other hand, by mature design and execution tools used in industrial processes and enterprise information management systems. We evaluate our approach by applying the DisCoveR miner to a large benchmark of real-life and synthetic event logs, measuring the size, density, separability, and constraint variability of mined models with and without grouping of activities. In earlier work, these measures have been shown to have a significant effect on the intrinsic cognitive load for users of declarative models, in particular DCR Graphs. We also evaluate the effect of prioritizing in particular the grouping of activities that model mutual exclusive choices. Our evaluation confirms that grouping of activities in general lowers the complexity on 3 of the 4 measures, while prioritizing choices in some cases makes the improvement slightly smaller.