Adaptive allocation of computation per sample

Develop an adaptive method that provisions the appropriate amount of computation for each batch of samples in GFlowNet training, and establish how such an approach should be implemented.

Background

The paper argues that the computational advantage of Particle GFlowNets depends on the relative costs of reward evaluation, policy evaluation, and trajectory sampling. It suggests that an ideal algorithm would adapt the computational budget to each batch of samples, but explicitly states that the implementation of such an approach remains unresolved.

References

With this in mind, as discussed in \Cref{sec:discussions}, we believe the ideal algorithm would adaptively provision the appropriate amount of computation for each batch of samples. How such an approach would have to be implemented, however, remains open.

— Particle GFlowNets: Rethinking Generative Marginalization Models  (2609.11538 - Silva et al., 10 Sep 2026) in Appendix, Section “Related works”