Rooted Absorbed Prefix Trajectory Balance with Submodular Replay for GFlowNet Training

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

Wang, X., Lu, W., & Wang, S. In International Conference on Machine Learning (ICML), 2026.
Paper