GFlowNet
A generative model framework proposed by Yoshua Bengio's team that models the generation process as a flow problem on a directed acyclic graph. It samples diverse terminal states with probability proportional to a reward signal, naturally avoiding mode collapse, and is suited to scenarios requiring diverse, high-quality outputs.
Timeline (last 90 days)
SNACK allows GFlowNet to be trained directly in string space and to explicitly handle node ordering symmetry.
Experiments show that compared with methods that operate directly on graphs, SNACK has a significant improvement in GFlowNet training throughput.
DOGS uses a GFlowNet sampler to generate design graphs, with generation probability proportional to the terminal reward.
The objective of GFlowNet training is to make the probability distribution of generated objects proportional to the reward function, rather than maximizing a single expected reward.
Generative Flow Networks (GFlowNets) were proposed by Yoshua Bengio's team in 2021.
GFlowNets model the generation process as a flow problem on a directed acyclic graph (DAG), training for flow conservation so that different terminal states are sampled with probability proportional to their reward.
The research team validated the method in two domains: tutoring dialogues and emotional support dialogues.
An arXiv paper proposes using Generative Flow Networks (GFlowNets) to generate diverse and high-quality synthetic expert dialogue data.
The paper emphasizes that the method generates high-quality data without directly replicating the training data.
On three outcome prediction tasks, classifiers trained on dialogues generated by GFlowNets provided a stronger training signal than models trained on data from other competing synthetic methods.
4 more timeline events
All Facts (14)
SNACK allows GFlowNet to be trained directly in string space and to explicitly handle node ordering symmetry.
50%UnverifiedExperiments show that compared with methods that operate directly on graphs, SNACK has a significant improvement in GFlowNet training throughput.
50%UnverifiedDOGS uses a GFlowNet sampler to generate design graphs, with generation probability proportional to the terminal reward.
50%UnverifiedThe objective of GFlowNet training is to make the probability distribution of generated objects proportional to the reward function, rather than maximizing a single expected reward.
50%UnverifiedGenerative Flow Networks (GFlowNets) were proposed by Yoshua Bengio's team in 2021.
50%UnverifiedGFlowNets model the generation process as a flow problem on a directed acyclic graph (DAG), training for flow conservation so that different terminal states are sampled with probability proportional to their reward.
50%UnverifiedThe research team validated the method in two domains: tutoring dialogues and emotional support dialogues.
50%UnverifiedAn arXiv paper proposes using Generative Flow Networks (GFlowNets) to generate diverse and high-quality synthetic expert dialogue data.
50%UnverifiedThe paper emphasizes that the method generates high-quality data without directly replicating the training data.
50%UnverifiedOn three outcome prediction tasks, classifiers trained on dialogues generated by GFlowNets provided a stronger training signal than models trained on data from other competing synthetic methods.
50%UnverifiedCompared to reinforcement learning baselines and end-to-end LLM baselines, the GFlowNet method achieves a better balance across three dimensions: fidelity, mode coverage, and authenticity.
50%UnverifiedGFlowNets can sample in proportion to the frequency with which each strategy appears in the training data, thereby preserving the diversity of the strategy distribution.
50%UnverifiedThe paper's approach is to train GFlowNets to generate the latent conversation structure rather than directly generating surface-level text.
50%UnverifiedThe latent structure is characterized by building a Gaussian mixture density over key interaction features, including confusion episode dynamics and the instructional balance of scaffolded guidance.
50%