Sparse Reward-Adaptive Generative Flow Networks

ICLR 2026 Conference Submission21256 Authors

Keywords: GFlowNet, sparse rewards, trajectory balance, generative models

TL;DR: We analyze and improve GFlowNet training in environments with sparse rewards

Abstract: Generative Flow Networks (GFlowNets) are an emerging class of algorithms for learning policies that sample objects according to an unnormalized reward distribution. While theoretically appealing, in practice, GFlowNets often suffer from training instabilities and mode collapse in environments with sparse rewards. These limit their applicability in a wide range of problems in which high-reward samples are valuable but sparse. In this paper, we identify and analyze three key challenges in training GFlowNets within sparse-reward environments and propose simple and targeted methods to mitigate each of them. Through extensive evaluation across various benchmark environments spanning both discrete and continuous problems, we demonstrate that our methods significantly improve training stability and policy quality, enabling GFlowNets to more reliably discover and exploit high-reward modes in challenging settings.

Primary Area: generative models

Submission Number: 21256