Researchers have studied the performance of a type of AI model called prediction cascades when faced with degraded input data. Prediction cascades route easy inputs through a lightweight model and difficult cases to a larger one, reducing energy consumption while maintaining accuracy. However, this design relies on the reliability of confidence-based routing decisions, which can be affected by input degradations such as static corruptions or sequential perturbations. The study found that these degradations can lead to three failure modes in prediction cascades: breaking the routing signal, degrading both models, and suppressing deferral while stabilizing predictions. This highlights the need for evaluating energy-efficient model cascades beyond clean accuracy and considering routing reliability under distribution shift. --- Why it matters: This research is important because it challenges the assumption that energy-efficient AI models can maintain high performance even when faced with degraded data. Understanding how prediction cascades behave in such scenarios can help engineers design more robust models that balance efficiency and accuracy. Source: https://arxiv.org/abs/2608.17711