Revisiting Non-Acyclic GFlowNets in Discrete Environments

Authors: Nikita Morozov, Ian Maksimov, Daniil Tiapkin, Sergey Samsonov

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In addition, we experimentally re-examine the concept of loss stability in nonacyclic GFlow Net training, as well as validate our own theoretical findings.
Researcher Affiliation Academia 1HSE University, Moscow, Russia 2CMAP CNRS Ecole polytechnique Institut Polytechnique de Paris, 91128, Palaiseau, France 3Universit e Paris-Saclay, CNRS, LMO, 91405, Orsay, France. Correspondence to: Nikita Morozov <EMAIL>.
Pseudocode No The paper describes methods and algorithms in paragraph form and through mathematical equations, but it does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Source code: github.com/Great Drake/non-acyclic-gfn.
Open Datasets Yes We consider two discrete environments for experimental evaluation: 1) a non-acyclic version of the hypergrid environment (Bengio et al., 2021) that was introduced in (Brunswic et al., 2024); 2) non-acyclic permutation generation environment from (Brunswic et al., 2024) with a harder variant of the reward function.
Dataset Splits No The paper focuses on generative models that sample objects from a distribution. It evaluates the models based on empirical distributions of generated samples (e.g., 'last 2 * 10^5 samples seen in training' or 'last 10^5 samples seen in training'), rather than using predefined training, validation, and test splits for an input dataset in a discriminative task.
Hardware Specification No This research was supported in part through computational resources of HPC facilities at HSE University (Kostenetskiy et al., 2021).
Software Dependencies No The paper mentions using the 'Adam optimizer' but does not specify its version or the versions of any other software libraries or programming languages used.
Experiment Setup Yes We use Adam optimizer with a learning rate of 10^-3 and a batch size of 16 (number of trajectories sampled at each training step). For log Zθ we use a larger learning rate of 10^-2... All models are trained until 2 * 10^6 trajectories are sampled... For SDB we set ε = 1.0 and η = 10^-3.