Stable GFlowNets with Probabilistic Guarantees

TL;DR

This paper introduces Stable GFlowNets, a new algorithm that stabilizes training and improves distributional fidelity by leveraging theoretical bounds relating loss to distribution accuracy.

Contribution

The paper provides a theoretical analysis of GFlowNet training stability and proposes a novel algorithm that ensures more reliable and faithful sampling.

Findings

Stable GFlowNets exhibit fewer training instabilities.

Theoretical bounds connect loss metrics to distribution fidelity.

Empirical results show improved training stability and accuracy.

Abstract

Generative Flow Networks (GFlowNets) learn to sample states proportional to an unnormalized reward. Despite their theoretical promise, practical training is often unstable, exhibiting severe loss spikes and mode collapse. To tackle this, we first assess the sensitivity of GFlowNet objectives, demonstrating that a small Total Variation (TV) distance between the learned and target distributions does not preclude unbounded training loss. Motivated by this mismatch, we establish converse guarantees by deriving loss-to-TV bounds that certify global fidelity from bounded trajectory balance losses. Lastly, we propose Stable GFlowNets, an algorithm that leverages our theoretical results to stabilize training, and empirically demonstrate improved training behavior and superior distributional fidelity.