QSTAR: Quantum Selective Transfer with Adaptive Routing
Abstract
Quantum transfer learning (QTL) is often evaluated by replacing a classical classifier with a fixed variational quantum head, but this hides a key question: when is the quantum branch actually useful? We propose QSTAR: Quantum Selective Transfer with Adaptive Routing, a selective QTL framework that keeps high-confidence classical predictions and routes only low-confidence samples to a fallback branch. Using a frozen ResNet18 backbone on Fashion-MNIST, we compare manually designed QTL heads, KetGPT-designed quantum heads, and parameter-matched classical baselines under a common data split and optimization schedule. Standard QTL heads reach at most 57.0% accuracy, while the strongest KetGPT head in the main filtered sweep reaches 78.5% accuracy and 0.785 F1-score. Although the strongest fixed classical head remains higher at 81.6%, selective routing gives the quantum branch a clearer role. On low-confidence samples, KetGPT #180 improves accuracy over a parameter-matched MLP fallback by 6.82, 4.31, and 3.03 percentage points at thresholds of 0.70, 0.80, and 0.90. At the full-system level, Adaptive KetGPT-QTL reaches 80.9% accuracy and 0.807 F1-score, outperforming the adaptive classical baseline. A separate compact-circuit ablation identifies KetGPT #160 as a stronger fixed-head candidate, reaching 81.9% accuracy with only 10 quantum parameters and 9 gates. These results suggest that architecture-searched quantum heads are most useful as targeted fallback branches for uncertain inputs rather than uniform replacements for classical classifiers.
Related
- Quantum State Routing and Perfect State Transfer on Signed Graphs under Environmental Noise
- Wireless Energy Transfer for Space Applications: A Review of Technologies, Mission Architectures, Energy Routing, and System Integration
- Tight Communication Bounds for Distributed Algorithms in the Quantum Routing Model
- Orbital transfer reveals material-selective structure in Rydberg excitons
- Energy Internet Routing using Quantum Optimization Algorithms
BibTeX
@misc{rehman2026qstar,
title = {{QSTAR: Quantum Selective Transfer with Adaptive Routing}},
author = {Saim Rehman and Nouhaila Innan and Muhammad Shafique},
howpublished = {arXiv},
year = {2026},
month = jul,
doi = {10.48550/arxiv.2607.21411},
eprint = {2607.21411},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2607.21411},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - QSTAR: Quantum Selective Transfer with Adaptive Routing
AU - Saim Rehman
AU - Nouhaila Innan
AU - Muhammad Shafique
PY - 2026
DA - 2026/07/23
PB - arXiv
DO - 10.48550/arxiv.2607.21411
UR - https://arxiv.org/abs/2607.21411
AB - Quantum transfer learning (QTL) is often evaluated by replacing a classical classifier with a fixed variational quantum head, but this hides a key question: when is the quantum branch actually useful? We propose QSTAR: Quantum Selective Transfer with Adaptive Routing, a selective QTL framework that keeps high-confidence classical predictions and routes only low-confidence samples to a fallback branch. Using a frozen ResNet18 backbone on Fashion-MNIST, we compare manually designed QTL heads, KetGPT-designed quantum heads, and parameter-matched classical baselines under a common data split and optimization schedule. Standard QTL heads reach at most 57.0% accuracy, while the strongest KetGPT head in the main filtered sweep reaches 78.5% accuracy and 0.785 F1-score. Although the strongest fixed classical head remains higher at 81.6%, selective routing gives the quantum branch a clearer role. On low-confidence samples, KetGPT #180 improves accuracy over a parameter-matched MLP fallback by 6.82, 4.31, and 3.03 percentage points at thresholds of 0.70, 0.80, and 0.90. At the full-system level, Adaptive KetGPT-QTL reaches 80.9% accuracy and 0.807 F1-score, outperforming the adaptive classical baseline. A separate compact-circuit ablation identifies KetGPT #160 as a stronger fixed-head candidate, reaching 81.9% accuracy with only 10 quantum parameters and 9 gates. These results suggest that architecture-searched quantum heads are most useful as targeted fallback branches for uncertain inputs rather than uniform replacements for classical classifiers.
ER -