Neural networks implicitly define a latent vector field on the data manifold, via autoencoding iterations🌀 This representation retains properties of the model, revealing memorization and generalization regimes, and characterizing distribution shifts 📜: arxiv.org/abs/2505.22785