Organizations: Department of Electrical and Electronic Engineering, University of Melbourne, Parkville, 3010, VIC, Australia. · CSIRO, Sydney, NSW, Australia.
Multi-task learning (MTL) improves generalization and data efficiency by jointly learning related tasks through shared representations. In the widely used hard-parameter-sharing setting, a shared backbone is combined with task-specific prediction heads. However, task-specific parameters can grow rapidly with the number of tasks. Therefore, designing multi-task heads that preserve task specialization while improving parameter efficiency remains a key challenge. In Quantum Machine Learning (QML), variational quantum circuits (VQCs) provide a compact mechanism for mapping classical data to quantum states residing in high-dimensional Hilbert spaces, enabling expressive representations within constrained parameter budgets. We propose a parameter-efficient quantum multi-task learning (QMTL) framework that replaces conventional task-specific linear heads with a fully quantum prediction head in a hybrid architecture. The model consists of a VQC with a shared, task-independent quantum encoding stage, followed by lightweight task-specific ansatz blocks enabling localized task adaptation while maintaining compact parameterization. Under a controlled and capacity-matched formulation where the shared representation dimension grows with the number of tasks, our parameter-scaling analysis demonstrates that a standard classical head exhibits quadratic growth, whereas the proposed quantum head parameter cost scales linearly. We evaluate QMTL on three multi-task benchmarks spanning natural language processing, medical imaging, and multimodal sarcasm detection, where we achieve performance comparable to, and in some cases exceeding, classical hard-parameter-sharing baselines while consistently outperforming existing hybrid quantum MTL models with substantially fewer head parameters. We further demonstrate QMTL's executability on noisy simulators and real quantum hardware, illustrating its feasibility.
Figures & tables
Figure 1: Proposed Quantum Circuit architecture for multi-task learning. An extracted feature vector Z is embedded into a Q -qubit register via a shared trainable state-preparation stage fQ(Z,Θ) . This embeds Z to a quantum state ∣ψΘ(Z)⟩ in superposition over computational basis states (a vector in a 2Q -dimensional complex Hilbert space). The register is then partitioned into T disjoint qubit sub-registers, Qt⊆{q0,…,qQ−1} , and task-specific parameterized blocks {fQt}t=1T are applied on each sub-register to produce T measurement outcomes {y^t}t=1T within one circuit execution. This enables parameter-efficient MTL by sharing a common trainable backbone while allowing task specialization through localized operations and readouts.
Figure 2: Total number of trainable parameters in our proposed architecture (shown in Fig. 1 ) compared against the simplest classical counterpart. We assume that the output dimension of each task is r to aid analysis in this section.
Figure 3: High-level overview of the experimental pipeline: Dataset → Backbone encoder → MTL head variant → Evaluation metrics
Dataset
Head variant
Head detail
Training method
GLUE (Frozen BERT backbone)
Classical MTL head
1 linear layer per task
Task-sampled MTL (with capped batches and LR scheduler) Classification: BCEWithLogitsLoss , CrossEntropyLoss , Regression: MSELoss
HQNN (4 or 10 qubits)
VQC ( Phukan et al., 2024 ) followed by 1 linear layer per task
QMTL (10 qubits)
QMTL head ( S1→8=1 & S9=2 )
CheXpert (DenseNet121 backbone)
Classical MTL head
1 linear layer per task
Parallel tasks MTL Masked loss for missing labels; Class-weighted focal loss (Appendix Eq. 24 ) 3-class train, binary evaluations
HQNN (4 or 10 qubits)
VQC ( Phukan et al., 2024 ) followed by 1 linear layer per task
QMTL (10 qubits)
QMTL head ( St=2 for all tasks)
Table 1: Summary of model variants, backbone configuration, and training procedures used across datasets. Detailed hyperparameters are provided in Appendix 12 .
Figure 4: Total trainable parameters in task-specific heads across datasets. QMTL can achieve more than 12× reduction compared to classical and 4-5 × compared to HQNN (when scaled to equivalent qubit count)
Figure 5: Multi-task performance on the GLUE benchmark across four MTL head variants
Figure 6: Multi-task performance on the CheXpert dataset across four MTL head variants
Figure 7: Multi-task performance on the Extended MUStARD dataset across four MTL head variants
Figure 8: Effect of increasing parameter cost on model performance. The blue linegraph represents the trend observed when varying the encoding layer depth L∈{2,3,4} , while keeping the task-specific subcircuit depth Lh=1 . The orange linegraph shows the trend when we vary the task-specific subcircuit depth Lh∈{1,2,3} and keep the encoding layer depth L=3 . A moderately deep encoding subcircuit coupled with shallow task-specific subcircuits provides the best performance, highlighting the best balance between expressivity of the shared quantum representation and parameter efficiency.
Figure 9: Effect of entanglement in the shared encoder on QMTL performance across CheXpert tasks. These results suggest that entanglement in the shared encoder can improve the multi-task performance of QMTL.
Figure 10: Multi-task performance of QMTL on the CheXpert dataset under increasing depolarizing noise levels in AerSimulator . Accuracy (top) and F1-score (bottom) are reported across CheXpert tasks using identical trained weights.
Figure 11: Multi-task performance of QMTL on IBM Quantum devices ( ibm_fez , ibm_boston ) compared to noise-free statevector simulation ( default.qubit ). Accuracy (top) and F1-score (bottom) are reported across the CheXpert tasks using the same trained weights. ibm_boston performs closer to the ideal simulation than ibm_fez
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Modality
# tasks
Task Type
# samples
Labeling scheme
GLUE Benchmark
Text (NLP)
9
Binary, Multiclass, Regression
≤300k per task
Exactly 1-labeled task per sample
CheXpert
Chest X-ray (image)
5
Binary
20 k total
≥1 labeled task per sample
Extended MUStARD
Multimodal (audio, text, visual)
5
Binary, Multiclass
690 total
Fully labeled
Appendix
Table 2: Summary of datasets considered in this paper for comparing our proposed QMTL architecture against baselines.
Dataset
Task
Metrics
CheXpert
Atelectasis †
Acc, P, R, F1
Cardiomegaly †
Acc, P, R, F1
Consolidation †
Acc, P, R, F1
Edema †
Acc, P, R, F1
Pleural Effusion †
Acc, P, R, F1
GLUE
COLA †
Matthews_corr, Acc
Appendix
Table 3: Summary of datasets, tasks, and evaluation metrics used across experiments. Symbols denote task types: † = Binary classification, ‡ = 3-class classification, ♣ = 9-class classification, △ = Regression. Abbreviations: Acc = Accuracy, P = Precision, R = Recall, F1 = F1-score, Matthews_corr = Matthews Correlation, Spearman = Spearman correlation, Pearson = Pearson correlation. For GLUE tasks, the primary metric is listed first.
Quantum machine learning (QML) is increasingly recognized as one of the most promising near-term applications of quantum computing, viewed as a next-frontier candidate beyond purely classical approaches. Hybrid quantum-classical models operationalize this potential by embedding a parameterized quantum circuit within a model where all other components remain classical-a design already applied to chemistry simulation, financial modeling, and image classification. However, their deployment in privacy-sensitive, multi-party settings is constrained by the need to avoid centralizing raw data and by the requirement that modern quantum circuits remain parameter-efficient to stay trainable at scale. In this paper, we address these constraints by evaluating federated learning (FL) as a means of combining a hybrid quantum-classical active party with a classical passive party, using Sherpa.ai's Blind Vertical FL (SBVFL) protocol to avoid centralizing raw data, while drastically reducing communication. We construct the split multiplicative periodic parity (SMPP) benchmark, following common QML design practice. On this task, our simulations show that SBVFL raises accuracy from 0.7227 to 0.8757 compared to local training, closely approaching non-private centralized accuracy, and that the hybrid quantum-classical model achieves this with substantially fewer trainable parameters than the classical neural networks and random forest alternatives. These results show that FL enables high-performing, privacy-preserving quantum-classical collaboration without centralizing raw data.
Carlos Cano, Daniel M. Jimenez-Gutierrez, Diego Sal +4
Quantum machine learning models for sequential data face scalability challenges with complex multivariate signals. We introduce the Hybrid Quantum Temporal Convolutional Network (HQTCN), which combines classical temporal windowing with a quantum convolutional neural network core. By applying a shared quantum circuit across temporal windows, HQTCN captures long-range dependencies while achieving significant parameter reduction. Evaluated on synthetic NARMA sequences and high-dimensional EEG time-series, HQTCN performs competitively with classical baselines on univariate data and outperforms all baselines on multivariate tasks. The model demonstrates particular strength under data-limited conditions, maintaining high performance with substantially fewer parameters than conventional approaches. These results establish HQTCN as a parameter-efficient approach for multivariate time-series analysis.
Junghoon Justin Park, Maria Pak, Sebin Lee +4
Seoul National University Seoul, Korea · Wells Fargo New York, USA · Brookhaven National Laboratory New York, USA
Rigorous empirical validation of quantum machine learning on natural language tasks remains scarce. We evaluate a 10-qubit hybrid quantum-classical variational circuit (2,148 parameters) for paraphrase detection across three benchmarks: MRPC, Quora Question Pairs (QQP), and adversarial PAWS. On QQP (n = 10 seeds), the circuit achieves 75.53% +/- 0.75%. accuracy, statistically outperforming parameter-matched classical baselines (DeepMLP: p = 0.015, Cohen's d = 1.20; F1: p < 0.001, d = 2.32) and surpassing DistilBERT-4bit with 31,191 fewer parameters. On MRPC the optimal 2-layer variant reaches 92% of BERT-base accuracy at 54,000 lower parameter cost. Circuit depth analysis reveals a dataset-depth scaling effect; entanglement analysis via the Meyer-Wallach measure identifies multi-qubit entanglement as the primary performance driver (r = 0.85 across four variants). Adversarial evaluation on PAWS reveals emergent robustness: 98.2% recall versus 81.6% classical (+16.6 pp, d = 1.24, p < 0.001), without adversarial training. These results constitute the first systematic parameter-matched multi-benchmark empirical validation of hybrid variational circuits for NLP. All results are from classical simulation; hardware validation is future work.
Farha Nausheen, Khandakar Ahmed, Farina Riaz
ISILC, Victoria University, Melbourne, VIC, Australia · CSIRO, Australian International Institute of Higher Education, Sydney, NSW, Australia