Information-Geometric Forward Policy Training in GFlowNets

This paper frames GFlowNet forward-policy optimization through the information geometry of its induced trajectory distribution. It identifies the trajectory family’s Fisher–Rao metric as the intrinsic first-order geometry and derives an exact decomposition of trajectory Fisher information into per-step conditional second moments. The authors distinguish exact, Monte Carlo, and structure-exploiting computational regimes, proposing graphical-model techniques such as marginalization, separator methods, and belief propagation as principled approximations. The abstract reports empirical comparisons between Riemannian and Euclidean optimization, but provides no datasets, numerical results, or implementation details.