EMBEr
Exploring Moment-Based Equivalence
for probabilistic program

15/04/2026

A piece of EMBEr at TACAS26

From April 12th to April 16th, TACAS (International Conference on Tools and Algorithms for the Construction and Analysis of Systems) was held in Turin as part of ETAPS (European Joint Conferences on Theory and Practice of Software).

Among the accepted papers, the first EMBEr-funded work was presented.

EMBEr focuses on testing moment-equivalence for probabilistic programs. A closely related problem is how moment information can be used effectively in core probabilistic programming tasks.

The paper “DeGAS: Gradient-based Optimization of Probabilistic Programs without Sampling” addresses this by extending a moment-matching approach for Gaussian mixtures, previously introduced for inference (see this), to optimization.

DeGAS relies on exact moment computation to build a differentiable, analytical approximation of the output distribution of a probabilistic program. This makes it possible to apply gradient-based optimization without sampling. The resulting closed-form representation also supports more complex objectives, such as constrained reachability probabilities. DeGAS is fully implemented in PyTorch, to leverage automatic differentiation and efficient GPU acceleration.

For a quick overview of DeGAS, you can look at our poster.
In alternative, you can read the full paper here or you can experiment with DeGAS here.