Abstraction Sampling (AS) is an extension of importance sampling inspired by the concept of abstractions in automated planning. An important component of Abstraction Sampling is the abstraction function that determines how nodes are grouped together into abstract states. Existing abstraction functions based on context-based heuristics such as RandCB have been proposed, but may not be good for many problems. In this paper, we demonstrate that the problem of finding an optimal abstraction function can be framed as a variance minimization optimization problem. We propose a new method for learning abstraction functions parameterized using graph neural networks. Furthermore, we introduce a novel algorithm, Dynamic Abstraction Sampling that is capable of competitive results with respect to existing abstraction functions.