Beyond Mode Collapse: Generative Flow Networks for Diverse Synthetic Expert Conversations
This arXiv preprint (2609.38359v1) proposes using Generative Flow Networks (GFlowNets) to sample diverse high-quality synthetic expert conversations by training on latent conversation structure with a Gaussian mixture density over interaction features. Evaluated on tutoring and emotional-support dialogue domains, the approach reportedly improves fidelity, mode coverage and authenticity versus RL and end-to-end LLM baselines without copying training data, and classifiers trained on the synthetic conversations yield stronger signals on three downstream outcome-prediction tasks.