Abstract
Flow Matching (FM) has established itself as the state-of-the-art for modeling complex conditional distributions, especially in prediction tasks like robotics and weather forecasting. However, its deployment in safety-critical settings is hindered by a critical extrapolation hazard: flow models yield plausible outputs even for off-manifold conditions due to smoothness biases, leading to silent failures indistinguishable from valid predictions.
In this work, we introduce Diverging Flows, a novel approach that enables a single model to simultaneously perform conditional generation and native extrapolation detection by structurally enforcing inefficient transport for off-manifold inputs.
We evaluate our method on synthetic manifolds, cross-domain style transfer, and weather temperature forecasting, demonstrating effective detection of extrapolations without compromising predictive fidelity or inference latency.
These results establish Diverging Flows as a robust solution for trustworthy flow models, paving the way for reliable deployment in domains such as medicine, robotics, and climate science.
Blogger's Review: The Diverging Flows method introduces a significant advancement in flow models by integrating extrapolation detection. This not only enhances model safety but also sets a foundation for reliable applications in critical areas. The challenge of achieving this without sacrificing performance showcases the researchers' profound skill in technological innovation.