Rooted Absorbed Prefix Trajectory Balance with Submodular Replay for GFlowNet Training

Abstract

Mode collapse remains a critical challenge in training Generative Flow Networks, especially when fine-tuning language models for molecular generation. We propose Rooted Absorbed Prefix Trajectory Balance, a training objective that strengthens early-stage learning signals, together with a submodular replay strategy that promotes diversity. The approach improves both sample quality and mode coverage on molecular generation tasks.

In brief

A new trajectory-balance objective and diversity-promoting replay strategy for stable GFlowNet training.