This paper introduces Attentive Implicit Representation Networks (AIR-Nets),\na simple, but highly effective architecture for 3D reconstruction from point\nclouds. Since representing 3D shapes in a local and modular fashion increases\ngeneralization and reconstruction quality, AIR-Nets encode an input point cloud\ninto a set of local latent vectors anchored in 3D space, which locally describe\nthe object's geometry, as well as a global latent description, enforcing global\nconsistency. Our model is the first grid-free, encoder-based approach that\nlocally describes an implicit function. The vector attention mechanism from\n[Zhao et al. 2020] serves as main point cloud processing module, and allows for\npermutation invariance and translation equivariance. When queried with a 3D\ncoordinate, our decoder gathers information from the global and nearby local\nlatent vectors in order to predict an occupancy value. Experiments on the\nShapeNet dataset show that AIR-Nets significantly outperform previous\nstate-of-the-art encoder-based, implicit shape learning methods and especially\ndominate in the sparse setting. Furthermore, our model generalizes well to the\nFAUST dataset in a zero-shot setting. Finally, since AIR-Nets use a sparse\nlatent representation and follow a simple operating scheme, the model offers\nseveral exiting avenues for future work. Our code is available at\nhttps://github.com/SimonGiebenhain/AIR-Nets.\n