A single calibration protects early-exit prediction sets

If a network chooses where to stop after seeing an input, can its prediction set still cover the true label at the promised rate? Calibrating exits separately does not answer this: selecting among individually valid sets can reduce coverage of the set actually returned.

Selection-Safe Conformal Early Exit calibrates one score across all exits so label-free stopping can preserve coverage of the true label.

The authors ask whether a network can stop at an input-dependent exit without losing the coverage promised by its prediction sets. Separately calibrated exits do not ensure this: selecting an exit based on the input can undermine coverage of the returned set.

Selection-Safe Conformal Early Exit calibrates one score across the full path and makes each earlier set contain the final calibrated set. In matched encoder tests, its macro selected coverage was 0.901 at a 0.90 target, versus 0.814 for naive separate-exit selection.

This supports the method’s answer under the tested conditions; the guarantee is marginal and depends on exchangeable calibration and test examples.

The guarantee is marginal and relies on exchangeable calibration and test examples and a model, scores, transformations, and exit schedule frozen before calibration. It does not guarantee coverage for each input type or under arbitrary distribution shift.