After some 20 years of progress, reservoir modeling still raises practical and theoretical challenges. Standard workflows are primarily built in a linear fashion (fault framework modeling, stratigraphic modeling, gridding, petrophysical modeling, upscaling, flow simulation and history matching). Modifying decisions in an early step in this workflow requires performing again all the dependent steps and the associated quality controls. Whereas robustness has significantly improved and makes it now possible to improve automation and running of multiple scenarios, we identify four main limitations in the current workflows and corresponding research axes: 1- Well correlations are most often deterministic in reservoir models, whereas they are affected by significant uncertainty. We need new ways to effectively sample this uncertainty by using sedimentological concepts and propagate it in existing time-to-depth conversion and gridding workflows. 2- Determination of the fault connectivity is often suboptimal: connectivity is decided from seismic fault sticks before stratigraphic modeling, whereas fault displacement is, with well test data, one of the most important arguments to decide about fault connectivity. We argue that evaluating fault displacement earlier in structural modeling workflows would most probably help choosing more realistic fault connectivity patterns right from the beginning, or stochastically sampling possible fault networks. 3- Gridding is a complex task and should ideally account for geological structures, facies and permeability fields and well geometry. The discretization of flow equations imposes additional gridding constraints to guarantee the performance and accuracy of flow simulation. Stair-step grids are a significant improvement to better handle some of these constraints, but locally flexible unstructured grids are the only way forward to appropriately integrate all available information while honoring discretization constraints. 4- Scale management has been a buzzword in reservoir modeling for years, but the current practice is still to take into account at best for two scales, and this after the reservoir gridding stage. We think that multiple scales should be used earlier on, and that connectivity and topological considerations should be used to ensure consistency between scales and between models and first-order dynamic information.