What Does Disagreement Mean? A Semantic Identifiability Frontier for Pairwise Preference Learning

Abstract

Pairwise comparison data are a standard supervision signal in preference learning, but disagreement in such data can reflect stable latent heterogeneity, ephemeral stochasticity, or non-binary responses hidden by a forced-choice interface. We study which of these semantic distinctions are identifiable from the feedback channel itself. We introduce a nonparametric latent semantic model with three querywise quantities: aggregate directional preference ${\mu }_q$, persistent disagreement ${\tau }_q$, and non-binary mass ${\alpha }_q$. Under the anonymous binary channel, we prove an exact semantic fiber theorem: ${\mu }_q$ is point-identified, whereas $({\tau }_q,{\alpha }_q)$ vary over an explicit feasible region compatible with the same binary law. Same-annotator repeats identify ${\tau }_q$ through a covariance identity, abstention-enabled feedback identifies ${\alpha }_q$ directly, and together they point-identify the full semantic triple. We also derive finite-sample estimators, nonasymptotic concentration bounds, and semantics-aware action-recovery guarantees, and validate the theory on theorem-matched synthetic data.

arXiv comment53 pages. Best Paper Award at the 2nd Workshop on Epistemic Intelligence in Machine Learning (EIML), ICML 2026