Large-scale pre-trained vision models such as ViT, CLIP, and SAM provide strong foundations for diverse vision tasks, motivating recent Mixture-of-Experts (MoE) approaches that combine multiple experts. However, existing methods often rely on static or implicit routing strategies, limiting adaptability to task semantics and input characteristics. We propose a task-adaptive vision expert routing framework based on competency learning guided by predictive uncertainty. We define expert competency as the relative reduction in predictive uncertainty induced by inter-expert interaction, and formulate expert routing as a learning problem driven by this signal. Our method uses task embeddings derived from textual descriptions to guide expert routing, refines expert features through cross-expert interaction, and aggregates them adaptively into a unified representation. By directly optimizing routing and feature composition using an uncertainty-based competency signal, the model learns how expert collaboration improves task-specific prediction reliability. Extensive experiments on diverse vision tasks demonstrate superior generalization performance and adaptive routing behavior aligned with task semantics.
Han et al. (Wed,) studied this question.