Quan-dorcet: Tournament-Based One-vs-One Quantum Classification for Robust Single-Shot Inference

ICLR 2026 Conference Submission17915 Authors

Keywords: quantum, classification, tournaments, qml, encoding

TL;DR: Using round-robin tournaments as an output encoding increases the likelihood of a single-shot from a QML model being accurate.

Abstract: Quantum machine learning (QML) promises powerful classification capabilities, but suffers from fragile output encodings and high sampling demands—especially in multiclass settings. Traditional schemes such as one-hot and binary encoding either produce interpretable outputs too rarely or require many shots to achieve reliable predictions. We propose a decision aggregation framework for quantum multiclass classification based on round-robin tournament scoring. Each output qubit represents a binary comparison between class pairs, and the final prediction is determined by majority wins—yielding a Condorcet-style winner when one exists. This structure improves both the resolvability and accuracy of single-shot predictions, outperforming standard encodings under few-shot conditions. Our method retains global entanglement while localizing decision tasks, enabling robust inference without sacrificing expressivity. Empirical results show that this approach achieves high accuracy and interpretability with significantly fewer measurements, suggesting a promising direction for practical quantum classifiers.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 17915