GFlowNets

GFlowNets (Generative Flow Networks) originated at Mila (Yoshua Bengio et al., 2021, "Flow Network based Generative Models", arXiv 2106.04399) as a foundational technique for learning to sample diverse, high-reward candidates in proportion to a reward — now used across drug, molecule, and materials discovery.

The in-window continuation "Adaptive Teachers for Amortized Samplers" (NeurIPS 2024) uses teacher-guided off-policy training to improve GFlowNet exploration (Emmanuel Bengio, Nikolay Malkin, Yoshua Bengio).