Sorting from Counterexamples

What the math gives to ML

The paper turns proper ranking prediction into geometric volume reduction: permutations are represented by order cells in a cube, while observed pairwise constraints carve out a convex consistency body. Querying the ranking induced by the body's centroid guarantees constant-factor elimination of feasible volume, unlike independently weighted pairwise predictions that can form Condorcet cycles. This suggests a neural preference-learning wrapper that maintains uncertainty over item scores or ranking hypotheses and chooses proper ranked-list queries by centroid rounding, with explicit robustness mechanisms for adversarial counterexamples. The main transferable asset is a principled active-query rule that preserves globally valid rankings while obtaining explicit information gain.