Rapidash: Scalable Molecular Modeling Through Controlled Equivariance Breaking

Molecular generation and property prediction have traditionally relied on models with strict equivariance, particularly preserving E(3) symmetries through equivariant models. However, recent advances in large-scale models suggest that unconstrained architectures, trained on extensive datasets, can implicitly learn these symmetries. In this work, we explore whether strict equivariance is required and present \texttt{Rapidash}, an architecture that allows breaking exact equivariance constraints. Using this architecture, we achieve state-of-the-art performance in molecular generation and property prediction, surpassing traditional equivariant models.