Studies of discrete languages emerging when neural agents communicate to\nsolve a joint task often look for evidence of compositional structure. This\nstems for the expectation that such a structure would allow languages to be\nacquired faster by the agents and enable them to generalize better. We argue\nthat these beneficial properties are only loosely connected to\ncompositionality. In two experiments, we demonstrate that, depending on the\ntask, non-compositional languages might show equal, or better, generalization\nperformance and acquisition speed than compositional ones. Further research in\nthe area should be clearer about what benefits are expected from\ncompositionality, and how the latter would lead to them.\n