Generating GFlowNets as You Wish with Diffusion Process

Yuxin Li, Wangbo Zhao, Dongwen Tang, jiyao liu, Xiahai Zhuang, Guang Li, Dianbo Liu, Yang You, Kai Wang

Keywords: GFlowNet, Parameter generation

Abstract: Generative Flow Networks (GFlowNets) are probabilistic samplers that learn stochastic policies to generate diverse sets of high-reward objects, which is essential in scientific discovery tasks. However, most existing GFlowNets necessitate training, becoming costly as the diversity of GFlowNets expands and trajectory lengths increase. To alleviate this problem, we propose a method to Generate high-performing GFlowNet parameters based on a given model structure, called GenFlowNet. Specifically, we first prepare an autoencoder to extract latent representations of GeFlowNet parameters and reconstruct them. Then, a structure encoder is trained alongside a conditional latent diffusion model to generate the target GFlowNet parameters based on the given structure information. To the best of our knowledge, it is the first exploration to generate parameters of a probabilistic sampler using the diffusion process. It enables us to obtain a new GFlowNet without training, effectively reducing the trial-and-error cost during GFlowNet development. Extensive experiments on diverse structures and tasks validate the superiority and generalizability of our method.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3398