Cheap Verifiers, Large Blind Spots: Measuring the Reliability Cost of Cost-Saving Cascades
Inference cascades use cheap models for most queries and frontier models as verifiers for hard cases; researchers found this self-improving loop is self-defeating due to blind spots and reliability costs. First, the verifier's blind spot—the fraction of student errors it accepts—grows with student capability (from 0.12 to 0.55) and shrinks with verifier cap…