GFlowNets get natural gradient training with Fisher-Rao geometry

GFlowNets are a powerful framework for amortized inference over discrete and mixed discrete-continuous objects, letting researchers sample from complex distributions using only an unnormalized reward function. But training the forward policy has traditionally relied on Euclidean gradient descent, which ignores the underlying geometry of the distribution being learned. In a new preprint, Yordan Raykov and Rodrigo Veiga propose an information-geometric alternative: treating the forward policy as an induced trajectory sampler and deriving natural gradient updates based on the Fisher-Rao metric of the trajectory family.

The authors prove that the trajectory Fisher information decomposes exactly into per-step conditional second moments. This decomposition reveals when temporal score interactions vanish (making updates cheap) and when dense couplings appear under shared parameterization. Based on this, they identify three computational regimes: problems with tractable exact Fisher information, settings where Monte Carlo estimators suffice, and structure-exploitable cases where target locality or factorization allows accurate approximations. For the last regime, they leverage graphical-model machinery — including exact marginalization, separator methods, and belief propagation — to approximate natural-gradient updates efficiently.

Empirically, the framework demonstrates faster convergence and more systematic exploration compared to Euclidean optimization on illustrative examples. The practical takeaway: the geometric structure of the target distribution can be baked directly into the optimization dynamics, giving practitioners a principled way to train GFlowNets in high-dimensional or combinatorially large spaces where naive gradient methods struggle. The paper also opens the door to scaling GFlowNets to larger problems where reward structure is sparse or factorized.