arXiv:2602.13061v2 Announce Type: replace 
Abstract: The ability of Flow Matching (FM) to model complex conditional distributions has established it as the state-of-the-art for prediction tasks (e.g., robotics, weather forecasting). However, deployment in safety-critical settings is hindered by a critical extrapolation hazard: driven by smoothness biases, flow models yield plausible outputs even for off-manifold conditions, resulting in 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 that it achieves 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.

تقدم دراسة حديثة مفهوم Diverging Flows، وهو نهج مبتكر في Flow Matching يعزز من توليد الشروط ويكتشف الاستقراء الأصلي في نماذج التعلم الآلي. تتناول هذه الطريقة المشكلة الحرجة المتمثلة في إنتاج نماذج التدفق لمخرجات مقبولة في ظروف خارج المنحنى، مما قد يؤدي إلى فشل صامت في التطبيقات الحرجة للسلامة مثل الروبوتات وعلوم المناخ.

Un estudio reciente presenta Diverging Flows, un enfoque novedoso en Flow Matching que mejora la generación condicional y detecta la extrapolación nativa en modelos de aprendizaje automático. Este método aborda el problema crítico de que los modelos de flujo producen resultados plausibles en condiciones fuera del manifold, lo que puede llevar a fallos silenciosos en aplicaciones críticas para la seguridad, como la robótica y la ciencia climática.

Une étude récente présente Diverging Flows, une approche novatrice dans le Flow Matching qui améliore la génération conditionnelle et détecte l'extrapolation native dans les modèles d'apprentissage automatique. Cette méthode aborde le problème critique des modèles de flux produisant des résultats plausibles dans des conditions hors-manifold, ce qui peut entraîner des échecs silencieux dans des applications critiques pour la sécurité telles que la robotique et la science climatique.

A recent study introduces Diverging Flows, a novel approach in Flow Matching that enhances conditional generation and detects native extrapolation in machine learning models. This method addresses the critical issue of flow models producing plausible outputs under off-manifold conditions, which can lead to silent failures in safety-critical applications such as robotics and climate science.