We present the Rao-Blackwellised reparameterisation gradient estimator and show its performance gains on a suite of probabilistic models.

We reduce training variance in equivariant generative models using a low-variance gradient estimator, improving stability and performance across molecular, crystal, and protein generation tasks.