Postdoc, Princeton University
2 papers at NeurIPS 2025
We introduce Gatekeeper, a novel loss function that calibrates smaller models in cascade setups to confidently handle easy tasks while deferring complex ones, significantly improving deferral performance across diverse architectures and tasks.
We decompose the gap between selective classifiers and the ideal oracle into five measurable sources, showing that only non-monotone scoring methods can reduce it and improve reliability.
