Generative Flow Networks

Also known as GFlowNet

Overview

Generative Flow Networks (GFlowNets) build objects through sequences of actions and learn to sample completed objects in proportion to their rewards, offering an alternative to finding only the single best solution. Common training methods include flow matching, detailed balance, and trajectory balance, with replay, policy-gradient updates, and learned action abstractions used to improve exploration and efficiency. Applications include molecule and crystal design, graph generation, language-model reasoning, and strategic decision-making. Current work seeks to scale these methods while preserving diversity and stable training, and to establish when training objectives reliably guarantee that the learned sampler matches the target reward distribution.

Latest papers

CardsList

64 papers7 papers in the last four weeks, down 36% on the four weeks before. 0.1% of all new papers.

  1. Action abstractions for amortized sampling

    Oct 19, 2024Oussama Boussif, Léna Néhale Ezzine, Joseph D Viviano +6Hierarchical RLGenerative Flow Networks

  2. Amortizing intractable inference in large language models

    Oct 6, 2023Edward J. Hu, Moksh Jain, Eric Elmoznino +4Amortized InferenceLLM Fine-Tuning

  3. Expected flow networks in stochastic environments and two-player zero-sum games

    Oct 4, 2023Marco Jiralerspong, Bilun Sun, Danilo Vucetic +4Game-Playing AgentsZero-Sum Games

  4. GFlowNets and variational inference

    Oct 2, 2022Esmeralda S. Whitammer, Salem Lahlou, Tristan Deleu +5Generative Flow NetworksRL for Generative Models