Calibration-Aware Uncertainty Cascades for Heterogeneous Model Collaboration

A new arXiv preprint proposes a routing method for combining multiple models that uses calibration-aware uncertainty estimates to decide when to escalate a query to a larger, more expensive model. The approach aims to avoid the rigidity of trained routers, which are tied to fixed cost or accuracy trade-offs, while still balancing predictive quality against inference cost. The work targets heterogeneous model collaboration settings where different models offer complementary strengths.