Representation-Agnostic Probabilistic Programming

This paper addresses a significant limitation in current probabilistic programming languages: the tight coupling of model representations with inference algorithms. By introducing a factor abstraction with five fundamental operations, the authors propose a universal interface that allows for the mixing of different representations (discrete tables, Gaussians, sample-based approaches) within a single framework. This is a crucial step towards enabling more flexible and expressive probabilistic models, particularly for complex hybrid models that current tools struggle with. The potential impact is significant, as it could lead to more efficient and accurate inference in a wider range of applications.