Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2610.08216

Quantum Entangled Multimodal Fusion Networks (QEMFN): Resource-Aware Hybrid Vision-Language Fusion via Trainable Entanglement

Srikar Alla, Ali Shiri Sichani, Chi-Ren Shyu

cs.AIquant-phSubmitted Oct 6, 2026• Announced Oct 7, 2026• 0 citations
arXivHTMLPDF

Abstract

Multimodal vision-language systems typically fuse image and text embeddings through classical operators such as concatenation, attention, bilinear pooling, or tensor interactions. We propose Quantum Entangled Multimodal Fusion Networks (QEMFN), a hybrid quantum-classical framework that introduces parameterized entanglement as a structured inductive bias for multimodal fusion. Pretrained visual and textual features are projected into compact latent spaces, encoded as angle-parameterized quantum states, processed through intra-modal and paired cross-modal entangling circuits, and measured to produce fused representations for retrieval. Under matched parameter budgets and identical frozen CLIP backbones, QEMFN outperforms classical fusion baselines on COCO-5k and Flickr30k, including multilayer perceptron, tensor fusion, FiLM, cross-attention, compact transformer, and a dequantized paired-topology analogue. An ablation suite isolates the quantum module's contribution from the surrounding classical projections, and quantum-centric analyses report Meyer-Wallach entangling capability, expressibility, gradient variance against barren-plateau bounds, and entropy-performance correlation under controls for training progress alongside an intervention study on the entangling component. QEMFN is executed under shot-based estimation, a noise-modeled fake backend, and a real superconducting device with zero-noise extrapolation. This work does not claim quantum computational advantage; the contribution is the framework together with a controlled empirical and quantum-centric evaluation that positions trainable entanglement as an interpretable, hardware-executable fusion mechanism at scales accessible on contemporary devices.

AI Summary

The paper asks whether a small variational quantum circuit can provide a useful inductive bias for vision–language fusion under matched trainable-parameter budgets. QEMFN projects frozen CLIP image and text features onto six qubits per modality, applies four layers of single-qubit rotations and intra- and cross-modal CZ couplings, and reads out 24 local and two-qubit observables. Classical projections and an output head are trained jointly with the circuit using a contrastive objective. Across ten seeds, QEMFN reports COCO-5k image-to-text Recall@1 of $68.9\pm0.4\%$, versus $68.2\pm0.4\%$ for a classical paired-topology analogue, $68.1\pm0.5\%$ for a compact transformer, and $67.1\pm0.4\%$ without cross-modal entanglement. Text-to-image Recall@1 is $48.4\pm0.4\%$; COCO-to-Flickr30k transfer reaches $60.8\pm0.4\%$. Ablations vary projection and circuit trainability, connectivity, and entropy regularization. Entropy–performance correlations remain positive after controlling for epoch or validation loss, at $0.56$ and $0.49$, respectively. Resource diagnostics include entangling capability, expressibility, and empirical gradient variance compared with barren-plateau reference bounds. Reported Recall@1 falls to $68.2\%$ with 4096-shot simulation and $63.8\%$ on superconducting hardware, recovering to $65.1\%$ with zero-noise extrapolation. The paper explicitly claims no computational advantage: these circuits remain classically simulable. Evaluation uses full pairwise joint scoring, requiring one circuit evaluation per candidate. Larger-scale retrieval and proposed top-$K$ reranking are not established; stronger classical surrogates and formal dequantizability analysis remain open.

Likely largely AI-generatedIncrementalUnclear writingPossible errorOverclaims

Central claim may not hold: The specified training score compares complete image–text pairs and has constant self-similarity, leaving the compatibility score used for the reported image–text retrieval undefined. The stated four-layer, 12-qubit Ry circuit has 48 trainable angles rather than 96, and its fixed CZ gates do not define the reported intervention of freezing entangling gates. After all cross-modal gates are removed, the specified pure product-input circuit must have zero visual–textual entanglement entropy, contradicting the reported normalized value of 0.31. The stated Hilbert–Schmidt expressibility formula gives 1−1/4096 for a 12-qubit pure state relative to the Haar-average density matrix, not the reported 0.09. The reported Bonferroni-corrected Wilcoxon thresholds below 0.001 and 0.01 are incompatible with conventional exact two-sided testing of ten seed pairs across nine comparisons.

Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.

Related papers

PQFA: Parallel Quantum Feature Augmentation of Fused Representations for Multimodal Classification

Mingzhu Wang, Yun Shang

Most multimodal learning methods improve how heterogeneous representations are aligned and fused, while post-fusion enhancement remains less explored. We propose Parallel Quantum Feature Augmentation (PQFA), a hybrid quantum-classical framework that applies multiple shallow variational quantum circuits to fused multimodal features. Text and image representations extracted by frozen RoBERTa and ViT encoders are processed through bidirectional cross-attention, attentive pooling, and adaptive gated fusion. The fused feature is then amplitude-encoded into parallel quantum circuits, whose measurement readouts are concatenated with the classical representation for prediction. We evaluate PQFA on MM-IMDb and N24News through controlled comparisons using the same encoders, fusion backbone, data splits, projection dimension, and augmentation output width. PQFA consistently outperforms both the fusion backbone without quantum augmentation and a width-matched MLP augmentation baseline, while using approximately 2.2K augmentation parameters compared with 24.0K for the MLP branch. Missing-modality experiments further show improved robustness when textual or visual inputs are incomplete, with particularly clear gains when the more informative textual modality is severely degraded. Controlled ablations and feature-space analyses indicate that the improvement cannot be reproduced by random feature mappings, increased classical width, or untrained quantum transformations. Quantum-state diagnostics additionally show stable predictive performance across the tested simulated noise levels and distinct branch-specific transformations of the encoded states. These results establish PQFA as an effective and parameter-efficient strategy for post-fusion augmentation in hybrid quantum-classical multimodal learning.

cs.LGCross-listedarXiv:2607.13466•Jul 16, 2026•1 citation

Hybrid Quantum–Classical Convolutional Networks for Robust Multimodal Pattern Recognition

Qingmei Li

Robust multimodal pattern recognition remains challenging under real-world conditions characterized by sensor degradation, distribution shift, and missing modalities. This study introduces a Hybrid Quantum–Classical Convolutional Network (Hybrid QC–CNN) that combines modality-specific convolutional encoders with a compact Parameterized Quantum Circuit (PQC) feature transformation and mask-aware fusion to stabilize inference under partial and corrupted evidence. Under matched-capacity baselines, the proposed model achieves 0.913 clean accuracy and 0.907 macro-F1 while delivering superior robustness across corruption severities. Averaged over severity levels 1–5, Hybrid QC–CNN attains 0.812 accuracy compared to 0.782 (multimodal transformer) and 0.771 (CNN late fusion), and improves worst-severity (level 5) accuracy to 0.701, exceeding the transformer (0.602) by +9.9 percentage points and late fusion (0.583) by +11.8 points. Robustness summarized by AURC reaches 0.823, surpassing the transformer (0.793) and late fusion (0.781). Reliability under degradation is also improved: ECE is 0.032 on clean data and 0.061 at severity 3, compared to 0.089 (transformer) and 0.094 (late fusion). Missing-modality stress tests further show stable degradation: with only one modality available, Hybrid QC–CNN achieves 0.793 (image-only), 0.742 (audio-only), and 0.764 (text-only), outperforming both transformer and late-fusion baselines across all single-modality conditions. Ablation removing the quantum layer reduces average corruption accuracy from 0.812 to 0.792 and worst-severity accuracy from 0.701 to 0.641, indicating that the PQC contributes meaningful robustness beyond mask-aware fusion and regularization. Collectively, results support hybrid quantum–classical feature transformation as a targeted mechanism for multimodal robustness and calibrated confidence.

Journal of Quantum Artificial Intelligence•May 1, 2026•0 citations

Quantum-enhanced multimodal fusion for robust and accurate fake news detection

Thulasi Bikku

The digital era has witnessed a significant rise in misinformation, underscoring the urgent need for effective tools to distinguish factual news from fabricated content.This study proposes a novel methodology for fake news detection that leverages quantum entanglement to facilitate multimodal fusion of textual, visual, and acoustic data.The quantum entanglement algorithm offers significant advantages in managing high-dimensional data by enabling efficient feature optimization through quantum computing encodings integrated with neural network architectures.The existing quantum circuits for text, visual, and audio would co-occur while witnessing the message from humans to machines with the machines acting as quantum computers integrated with neural networks, specifically designed for fake news detection.The results generated from our method demonstrate significantly improved fake news detection accuracy, and increased accuracy in the noise simulation, and the system is resilient to adversarial methods, all in contrast to typical methods.The proposed Quantum Encoding with Multimodal Fusion (QEMF) framework surpasses existing approaches by offering as it suggests promising future directions to tackle fake news across the web.The system, in a scientific manner, performs a variety of strict pre-processing techniques to all textual data such as tokenization, stemming, and lemmatization, along with sophisticated image pre-processing to all visual data.It uses the latest extracting features, i.e., Glove for text embedding, and conventional convolutional networks, like VGG16, for visual data.The feature representation is significantly enriched through quantum encoding, while the capability of QCNNs to identify the most salient and discriminative features enhances accuracy and ensures robustness against noise and adversarial interference.Furthermore, the system's real-time detection capability and its scalability position it as a powerful tool in combating misinformation within the evolving bio-informational ecosystem.

Sigma Journal of Engineering and Natural Sciences – Sigma Mühendislik ve Fen Bilimleri Dergisi•Jan 1, 2025•3 citations

Expressive and Scalable Quantum Fusion for Multimodal Learning

Tuyen Nguyen, Hoang, Trong Nghia, Phi Le Nguyen +2 more

The aim of this paper is to introduce a quantum fusion mechanism for multimodal learning and to establish its theoretical and empirical potential. The proposed method, called the Quantum Fusion Layer (QFL), replaces classical fusion schemes with a hybrid quantum-classical procedure that uses parameterized quantum circuits to learn entangled feature interactions without requiring exponential parameter growth. Supported by quantum signal processing principles, the quantum component efficiently represents high-order polynomial interactions across modalities with linear parameter scaling, and we provide a separation example between QFL and low-rank tensor-based methods that highlights potential quantum query advantages. In simulation, QFL consistently outperforms strong classical baselines on small but diverse multimodal tasks, with particularly marked improvements in high-modality regimes. These results suggest that QFL offers a fundamentally new and scalable approach to multimodal fusion that merits deeper exploration on larger systems.

arXiv:2510.06938•Oct 8, 2025•3 citations

A Hybrid Quantum Re-Uploading Framework for Multimodal Emotion Recognition

Sai Siva Naga Chandra Pathuru, S. Sri Harsha

Multimodal emotion recognition can benefit from combining complementary information from facial images, speech, and text.However, integrating heterogeneous modalities obtained from independently collected datasets introduces challenges in data alignment and label consistency.This study presents a hybrid quantum-classical framework that combines modality-specific classical encoders with a quantum re-uploading variational classifier for six-class emotion recognition.Facial images, speech samples, and textual samples were obtained from separate publicly available datasets and aligned at the emotion-label level to construct a unified multimodal experimental setting.Image, speech, and text encoders generated modality-specific representations that were fused through a classical neural network and projected into a six-dimensional quantum input space.A six-qubit quantum circuit with three re-uploading layers repeatedly encoded the fused features before six-class classification.On a test set of 6,347 samples, the classical Image-Speech-Text fusion model achieved 73.94% accuracy and a 70.64% macro F1-score, whereas the proposed quantum reuploading model achieved 72.57% accuracy and a 68.76% macro F1-score.The quantum model used 8,706 trainable parameters compared with 15,440,006 for the classical multimodal model, representing a 99.94% reduction in trainable parameters.The results indicate that quantum re-uploading can provide competitive performance for multimodal emotion recognition while highlighting the limitations of label-level alignment across independently collected datasets.

International Journal of Drug Delivery Technology•Sep 16, 2026•0 citations

QiNN-QJ: A Quantum-inspired Neural Network with Quantum Jump for Multimodal Sentiment Analysis

Yiwei Chen, Kehuan Yan, Yu Pan +1 more

Quantum theory provides non-classical principles, such as superposition and entanglement, that inspires promising paradigms in machine learning. However, most existing quantum-inspired fusion models rely solely on unitary or unitary-like transformations to generate quantum entanglement. While theoretically expressive, such approaches often suffer from training instability and limited generalizability. In this work, we propose a Quantum-inspired Neural Network with Quantum Jump (QiNN-QJ) for multimodal entanglement modelling. Each modality is firstly encoded as a quantum pure state, after which a differentiable module simulating the QJ operator transforms the separable product state into the entangled representation. By jointly learning Hamiltonian and Lindblad operators, QiNN-QJ generates controllable cross-modal entanglement among modalities with dissipative dynamics, where structured stochasticity and steady-state attractor properties serve to stabilize training and constrain entanglement shaping. The resulting entangled states are projected onto trainable measurement vectors to produce predictions. In addition to achieving superior performance over the state-of-the-art models on benchmark datasets, including CMU-MOSI, CMU-MOSEI, and CH-SIMS, QiNN-QJ facilitates enhanced post-hoc interpretability through von-Neumann entanglement entropy. This work establishes a principled framework for entangled multimodal fusion and paves the way for quantum-inspired approaches in modelling complex cross-modal correlations.

arXiv:2510.27091•Oct 31, 2025•0 citations
Multimodal vision-language systems typically fuse image and text embeddings through classical operators such as concatenation, attention, bilinear pooling, or tensor interactions. We propose Quantum Entangled Multimodal Fusion Networks (QEMFN), a hybrid quantum-classical framework that introduces parameterized entanglement as a structured inductive bias for multimodal fusion. Pretrained visual and textual features are projected into compact latent spaces, encoded as angle-parameterized quantum states, processed through intra-modal and paired cross-modal entangling circuits, and measured to produce fused representations for retrieval. Under matched parameter budgets and identical frozen CLIP backbones, QEMFN outperforms classical fusion baselines on COCO-5k and Flickr30k, including multilayer perceptron, tensor fusion, FiLM, cross-attention, compact transformer, and a dequantized paired-topology analogue. An ablation suite isolates the quantum module's contribution from the surrounding classical projections, and quantum-centric analyses report Meyer-Wallach entangling capability, expressibility, gradient variance against barren-plateau bounds, and entropy-performance correlation under controls for training progress alongside an intervention study on the entangling component. QEMFN is executed under shot-based estimation, a noise-modeled fake backend, and a real superconducting device with zero-noise extrapolation. This work does not claim quantum computational advantage; the contribution is the framework together with a controlled empirical and quantum-centric evaluation that positions trainable entanglement as an interpretable, hardware-executable fusion mechanism at scales accessible on contemporary devices.
The paper asks whether a small variational quantum circuit can provide a useful inductive bias for vision–language fusion under matched trainable-parameter budgets. QEMFN projects frozen CLIP image and text features onto six qubits per modality, applies four layers of single-qubit rotations and intra- and cross-modal CZ couplings, and reads out 24 local and two-qubit observables. Classical projections and an output head are trained jointly with the circuit using a contrastive objective.
Across ten seeds, QEMFN reports COCO-5k image-to-text Recall@1 of 68.9±0.4%68.9\pm0.4\%68.9±0.4%, versus 68.2±0.4%68.2\pm0.4\%68.2±0.4% for a classical paired-topology analogue, 68.1±0.5%68.1\pm0.5\%68.1±0.5% for a compact transformer, and 67.1±0.4%67.1\pm0.4\%67.1±0.4% without cross-modal entanglement. Text-to-image Recall@1 is 48.4±0.4%48.4\pm0.4\%48.4±0.4%; COCO-to-Flickr30k transfer reaches 60.8±0.4%60.8\pm0.4\%60.8±0.4%. Ablations vary projection and circuit trainability, connectivity, and entropy regularization. Entropy–performance correlations remain positive after controlling for epoch or validation loss, at 0.560.560.56 and 0.490.490.49, respectively. Resource diagnostics include entangling capability, expressibility, and empirical gradient variance compared with barren-plateau reference bounds.
Reported Recall@1 falls to 68.2%68.2\%68.2% with 4096-shot simulation and 63.8%63.8\%63.8% on superconducting hardware, recovering to 65.1%65.1\%65.1% with zero-noise extrapolation. The paper explicitly claims no computational advantage: these circuits remain classically simulable. Evaluation uses full pairwise joint scoring, requiring one circuit evaluation per candidate. Larger-scale retrieval and proposed top-KKK reranking are not established; stronger classical surrogates and formal dequantizability analysis remain open.
Most multimodal learning methods improve how heterogeneous representations are aligned and fused, while post-fusion enhancement remains less explored. We propose Parallel Quantum Feature Augmentation (PQFA), a hybrid quantum-classical framework that applies multiple shallow variational quantum circuits to fused multimodal features. Text and image representations extracted by frozen RoBERTa and ViT encoders are processed through bidirectional cross-attention, attentive pooling, and adaptive gated fusion. The fused feature is then amplitude-encoded into parallel quantum circuits, whose measurement readouts are concatenated with the classical representation for prediction. We evaluate PQFA on MM-IMDb and N24News through controlled comparisons using the same encoders, fusion backbone, data splits, projection dimension, and augmentation output width. PQFA consistently outperforms both the fusion backbone without quantum augmentation and a width-matched MLP augmentation baseline, while using approximately 2.2K augmentation parameters compared with 24.0K for the MLP branch. Missing-modality experiments further show improved robustness when textual or visual inputs are incomplete, with particularly clear gains when the more informative textual modality is severely degraded. Controlled ablations and feature-space analyses indicate that the improvement cannot be reproduced by random feature mappings, increased classical width, or untrained quantum transformations. Quantum-state diagnostics additionally show stable predictive performance across the tested simulated noise levels and distinct branch-specific transformations of the encoded states. These results establish PQFA as an effective and parameter-efficient strategy for post-fusion augmentation in hybrid quantum-classical multimodal learning.
Robust multimodal pattern recognition remains challenging under real-world conditions characterized by sensor degradation, distribution shift, and missing modalities. This study introduces a Hybrid Quantum–Classical Convolutional Network (Hybrid QC–CNN) that combines modality-specific convolutional encoders with a compact Parameterized Quantum Circuit (PQC) feature transformation and mask-aware fusion to stabilize inference under partial and corrupted evidence. Under matched-capacity baselines, the proposed model achieves 0.913 clean accuracy and 0.907 macro-F1 while delivering superior robustness across corruption severities. Averaged over severity levels 1–5, Hybrid QC–CNN attains 0.812 accuracy compared to 0.782 (multimodal transformer) and 0.771 (CNN late fusion), and improves worst-severity (level 5) accuracy to 0.701, exceeding the transformer (0.602) by +9.9 percentage points and late fusion (0.583) by +11.8 points. Robustness summarized by AURC reaches 0.823, surpassing the transformer (0.793) and late fusion (0.781). Reliability under degradation is also improved: ECE is 0.032 on clean data and 0.061 at severity 3, compared to 0.089 (transformer) and 0.094 (late fusion). Missing-modality stress tests further show stable degradation: with only one modality available, Hybrid QC–CNN achieves 0.793 (image-only), 0.742 (audio-only), and 0.764 (text-only), outperforming both transformer and late-fusion baselines across all single-modality conditions. Ablation removing the quantum layer reduces average corruption accuracy from 0.812 to 0.792 and worst-severity accuracy from 0.701 to 0.641, indicating that the PQC contributes meaningful robustness beyond mask-aware fusion and regularization. Collectively, results support hybrid quantum–classical feature transformation as a targeted mechanism for multimodal robustness and calibrated confidence.
The digital era has witnessed a significant rise in misinformation, underscoring the urgent need for effective tools to distinguish factual news from fabricated content.This study proposes a novel methodology for fake news detection that leverages quantum entanglement to facilitate multimodal fusion of textual, visual, and acoustic data.The quantum entanglement algorithm offers significant advantages in managing high-dimensional data by enabling efficient feature optimization through quantum computing encodings integrated with neural network architectures.The existing quantum circuits for text, visual, and audio would co-occur while witnessing the message from humans to machines with the machines acting as quantum computers integrated with neural networks, specifically designed for fake news detection.The results generated from our method demonstrate significantly improved fake news detection accuracy, and increased accuracy in the noise simulation, and the system is resilient to adversarial methods, all in contrast to typical methods.The proposed Quantum Encoding with Multimodal Fusion (QEMF) framework surpasses existing approaches by offering as it suggests promising future directions to tackle fake news across the web.The system, in a scientific manner, performs a variety of strict pre-processing techniques to all textual data such as tokenization, stemming, and lemmatization, along with sophisticated image pre-processing to all visual data.It uses the latest extracting features, i.e., Glove for text embedding, and conventional convolutional networks, like VGG16, for visual data.The feature representation is significantly enriched through quantum encoding, while the capability of QCNNs to identify the most salient and discriminative features enhances accuracy and ensures robustness against noise and adversarial interference.Furthermore, the system's real-time detection capability and its scalability position it as a powerful tool in combating misinformation within the evolving bio-informational ecosystem.
The aim of this paper is to introduce a quantum fusion mechanism for multimodal learning and to establish its theoretical and empirical potential. The proposed method, called the Quantum Fusion Layer (QFL), replaces classical fusion schemes with a hybrid quantum-classical procedure that uses parameterized quantum circuits to learn entangled feature interactions without requiring exponential parameter growth. Supported by quantum signal processing principles, the quantum component efficiently represents high-order polynomial interactions across modalities with linear parameter scaling, and we provide a separation example between QFL and low-rank tensor-based methods that highlights potential quantum query advantages. In simulation, QFL consistently outperforms strong classical baselines on small but diverse multimodal tasks, with particularly marked improvements in high-modality regimes. These results suggest that QFL offers a fundamentally new and scalable approach to multimodal fusion that merits deeper exploration on larger systems.
Multimodal emotion recognition can benefit from combining complementary information from facial images, speech, and text.However, integrating heterogeneous modalities obtained from independently collected datasets introduces challenges in data alignment and label consistency.This study presents a hybrid quantum-classical framework that combines modality-specific classical encoders with a quantum re-uploading variational classifier for six-class emotion recognition.Facial images, speech samples, and textual samples were obtained from separate publicly available datasets and aligned at the emotion-label level to construct a unified multimodal experimental setting.Image, speech, and text encoders generated modality-specific representations that were fused through a classical neural network and projected into a six-dimensional quantum input space.A six-qubit quantum circuit with three re-uploading layers repeatedly encoded the fused features before six-class classification.On a test set of 6,347 samples, the classical Image-Speech-Text fusion model achieved 73.94% accuracy and a 70.64% macro F1-score, whereas the proposed quantum reuploading model achieved 72.57% accuracy and a 68.76% macro F1-score.The quantum model used 8,706 trainable parameters compared with 15,440,006 for the classical multimodal model, representing a 99.94% reduction in trainable parameters.The results indicate that quantum re-uploading can provide competitive performance for multimodal emotion recognition while highlighting the limitations of label-level alignment across independently collected datasets.
Quantum theory provides non-classical principles, such as superposition and entanglement, that inspires promising paradigms in machine learning. However, most existing quantum-inspired fusion models rely solely on unitary or unitary-like transformations to generate quantum entanglement. While theoretically expressive, such approaches often suffer from training instability and limited generalizability. In this work, we propose a Quantum-inspired Neural Network with Quantum Jump (QiNN-QJ) for multimodal entanglement modelling. Each modality is firstly encoded as a quantum pure state, after which a differentiable module simulating the QJ operator transforms the separable product state into the entangled representation. By jointly learning Hamiltonian and Lindblad operators, QiNN-QJ generates controllable cross-modal entanglement among modalities with dissipative dynamics, where structured stochasticity and steady-state attractor properties serve to stabilize training and constrain entanglement shaping. The resulting entangled states are projected onto trainable measurement vectors to produce predictions. In addition to achieving superior performance over the state-of-the-art models on benchmark datasets, including CMU-MOSI, CMU-MOSEI, and CH-SIMS, QiNN-QJ facilitates enhanced post-hoc interpretability through von-Neumann entanglement entropy. This work establishes a principled framework for entangled multimodal fusion and paves the way for quantum-inspired approaches in modelling complex cross-modal correlations.