$\alpha$-GFN: Generalized Mixing in GFlowNets for Better Exploration-Exploitation Trade-off

Lin Chen, Samuel Drapeau, Fanghao Shao, Xuekai Zhu, Bo Xue, Yunchong Song, Mathieu Lauriere, Zhouhan Lin

Keywords: GFlowNet, Markov chain, Generative Models

TL;DR: We propose α-GFNs: flexible forward–backward policy mixing with a parameter α that controls exploration–exploitation, grounded in an extended GFlowNet–Markov chain connection.

Abstract:

Standard Generative Flow Network (GFlowNet) training implicitly assigns equal weights to the forward and backward policies, a consequence of the flow-matching view that constrains the exploration–exploitation dynamics.
Extending the connection between GFlowNets and Markov chains, we show that this equal weighting arises from a theoretical equivalence between GFlowNet objectives and Markov chain reversibility.
Building on this, we introduce $\boldsymbol{\alpha}\textbf{-GFNs}$, which generalize standard GFlowNet training from strictly balanced flows to imbalanced flows by mixing the forward and backward policies with a hyperparameter $\alpha$ in the training objectives. Through the link to reversibility, we further establish that such objectives converge to unique flows. This generalization provides a richer exploration–exploitation trade-off and, in some settings, coarse control over trajectory lengths. We also propose a simple scheduling algorithm to combine the strengths of different $\alpha$ values. Experiments on Set Generation, Bit Sequence Generation, and Molecule Generation demonstrate consistent performance gains and highlight the benefits of $\alpha$ tuning.

Primary Area: generative models

Submission Number: 23558