Why Ask One When You Can Ask $k$? Two-Stage Learning-to-Defer to the Top-$k$ Experts

Although existing Learning-to-Defer (L2D) frameworks support multiple experts, they allocate each query to a single expert, limiting their ability to leverage collective expertise in complex decision-making scenarios. To address this, we introduce the first framework for Top-$k$ Learning-to-Defer, enabling systems to defer each query to the $k$ most cost-effective experts. Our formulation strictly generalizes classical two-stage L2D by supporting multi-expert deferral-a capability absent in prior work. We further propose Top-$k(x)$ Learning-to-Defer, an adaptive extension that learns the optimal number of experts per query based on input complexity, expert quality, and consultation cost. We introduce a novel surrogate loss that is Bayes-consistent, $(\mathcal{R}, \mathcal{G})$-consistent, and independent of the cardinality parameter $k$, enabling efficient reuse across different values of $k$. We show that classical model cascades arise as a special case of our method, situating our framework as a strict generalization of both selective deferral and cascaded inference. Experiments on classification and regression demonstrate that Top-$k$ and Top-$k(x)$ yield improved accuracy--cost trade-offs, establishing a new direction for multi-expert deferral in Learning-to-Defer.