Native Extrapolation Awareness in Flow-Based Conditional Generation

arXiv cs.AI · 6 hours ago

Researchers introduced Diverging Flows, a method that enables flow-matching models to detect when input conditions fall outside their training data distribution while still generating predictions. The approach structures the model to produce inefficient transport for off-manifold inputs, allowing simultaneous conditional generation and extrapolation detection. The method was tested on synthetic manifolds, style transfer, and weather forecasting, maintaining prediction quality and inference speed while identifying when conditions are out-of-distribution.