GFlowNets generate varied synthetic tutoring and support conversations

Synthetic conversations can help post-train language models for adaptive applications, where examples should reflect varied expert strategies and decisions. The paper says direct prompting and conditioning on intended uses tend to produce low-diversity data concentrated around dominant patterns. Its authors propose Generative Flow Networks, or GFlowNets, as an alternative for producing synthetic expert conversations. The study examines tutoring and emotional-support applications.

Rather than generating only the words of a dialogue, the method trains GFlowNets to create latent conversation structure. It represents key interaction features with a Gaussian mixture density. The examples named in the paper include how confusion develops during an exchange and the balance of scaffolding directives. The model then samples expert strategies according to how prevalent they are in the training data, a design intended to support coverage beyond the most common conversational patterns.

The researchers report results in two structurally different domains: tutoring and emotional-support dialogues. They say the GFlowNet approach offers a stronger balance of fidelity, coverage of distinct modes and authenticity than reinforcement-learning and end-to-end language-model baselines. The abstract also says the generated conversations do not copy training data. It does not provide numerical scores in the summary, so the reported comparison is qualitative there.

The team also tested whether the generated conversations could help with downstream prediction. Across three outcome-prediction tasks, classifiers trained on GFlowNet-generated conversations provided a stronger training signal than those trained using competing synthesis baselines. For builders, the reported results suggest that synthetic-data evaluation can consider both the range of strategies represented and usefulness on downstream tasks, not just the production of plausible dialogue. The paper’s findings are specific to its two studied domains and three prediction tasks.