Optimal rejuvenation scheduling for persistent Gibbs samplers

Determine how to optimally decide when to rejuvenate the persistent Gibbs sampler used by Particle GFlowNets.

Background

Particle GFlowNets maintain a persistent Gibbs sampler during training, and the paper uses a Gelman–Rubin statistic threshold to trigger rejuvenation, meaning that the particle states are refreshed with independent samples from the current forward policy. Although the experiments show that this heuristic improves exploration and accelerates convergence, the paper does not establish an optimal decision rule or threshold for scheduling such refreshes.