Stable GFlowNets with Probabilistic Guarantees

arXiv cs.LG / 5/5/2026

📰 NewsIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper analyzes why Generative Flow Networks (GFlowNets) can be practically unstable, showing that even small total variation (TV) distance between learned and target distributions does not necessarily prevent training loss from diverging.
It derives “converse” theoretical guarantees by establishing bounds that relate limited trajectory-balance loss to global fidelity, effectively turning bounded training losses into distributional guarantees.
Building on these results, the authors introduce Stable GFlowNets, a new training algorithm intended to reduce severe loss spikes and mitigate mode collapse.
Experiments indicate that Stable GFlowNets improves both training stability and distributional fidelity compared with prior approaches.
Overall, the work provides a clearer theoretical foundation for GFlowNet training and a practical method to make learning behavior more reliable.

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.