Beyond Category Assignments: Learning Sample-Prototype Dynamics for Generalized Category Discovery

Generalized Category Discovery (GCD) aims to recognize known categories while discovering novel categories from partially labeled data. Existing methods often exploit predicted sample-category relations as pseudo-label supervision for representation learning. However, such supervision does not fully exploit the intermediate geometric structure between samples and candidate prototypes, and relies on uncertain assignments that may reinforce early errors. We introduce a continuous perspective on GCD representation learning by expanding categorical assignments into sample-prototype transport, providing supervision on how representations evolve toward candidate prototypes. Based on this perspective, we propose **DynGCD**, a plug-and-play framework that models continuous sample-prototype transport through prototype-conditioned flow matching while jointly learning uncertain sample-prototype couplings. The couplings preserve multiple plausible category relations and determine which candidate transports contribute to flow learning, while the learned transport consistency and evolving representations in turn refine the couplings. This mutual refinement allows category relations and transport dynamics to progressively co-evolve during training, reducing reliance on fixed categorical pseudo-targets. DynGCD requires only sample representations and category prototypes during training, making it readily applicable to existing GCD frameworks while preserving their native inference procedures. Experiments on generic and fine-grained benchmarks across *parametric*, *non-parametric*, and *vision-language* GCD paradigms demonstrate consistent improvements, with particularly strong gains on novel categories.