Revisiting Non-Acyclic GFlowNets in Discrete Environments
Nikita Morozov, Ian Maksimov, Daniil Tiapkin, Sergey Samsonov · 2025
Read Revisiting Non-Acyclic GFlowNets in Discrete Environments by Nikita Morozov, Ian Maksimov, Daniil Tiapkin, Sergey Samsonov (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-07735.txt.
Description
Nikita Morozov, Ian Maksimov, Daniil Tiapkin, and Sergey Samsonov revisited the theory of GFlowNets in 2025 when their sampling graph contains cycles. These generative models learn to sample objects with a target probability by walking a directed graph. Standard theory needs that graph to have no cycles, which limits designs.
The authors present a simpler theory for the cyclic case in discrete settings. New results cover training with fixed backward policies, the nature of flow functions, and links to entropy-regularised reinforcement learning. Each result generalises its acyclic counterpart. Experiments recheck how stable training losses are without acyclicity and confirm the new findings. The work widens where these samplers can be applied.
Catalog record
| CREATOR | Nikita Morozov, Ian Maksimov, Daniil Tiapkin, Sergey Samsonov |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 108 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-07735