Abstract Data fusion—the synergistic integration of information from multiple sources—has evolved from a statistical post-processing step into a foundational paradigm for modern artificial intelligence. This review highlights recent breakthroughs in deep multimodal fusion, uncertainty-aware architectures, and cross-modal generative models. We discuss how transformer-based attention mechanisms and graph neural networks have enabled scalable fusion of unstructured (text, imagery, audio) and structured (sensor, genomic, financial) data. Key advances include federated fusion for privacy-preserving analytics, physics-informed fusion for scientific discovery, and self-supervised fusion that reduces annotation burdens. We conclude by outlining open challenges: dynamic reconfiguration, causal fusion, and the ethical implications of fused decision-making.

1. Introduction The proliferation of sensing platforms—from Earth-observation satellites to wearable biosensors—has created an unprecedented deluge of heterogeneous data. Yet raw data streams are rarely informative in isolation. Data fusion, defined as the framework for combining data from multiple sources to achieve improved accuracy, robustness, and interpretability than any single source alone, has become central to autonomous navigation, medical diagnostics, climate modeling, and defense systems. Early fusion methods (e.g., Kalman filters, Dempster-Shafer theory) assumed linearity and known noise distributions. The deep learning era, however, has shifted the focus toward learning fusion representations directly from raw signals. This article synthesizes research from 2022–2025, emphasizing architectural innovations and their practical implications.

2. Recent breakthroughs in fusion architectures

2.1 Transformer-based cross-modal attention The introduction of cross-attention mechanisms in transformer models has revolutionized fusion of sequential and non-sequential modalities. ThePerceiverarchitecture (Jaegle et al., 2021) demonstrated that a latent array of fixed size can iteratively attend to arbitrary input modalities, enabling fusion of point clouds, audio, and images without modality-specific encoders. More recently,ImageBind(Girdhar et al., 2023) from Meta AI showed that binding six modalities (images, text, audio, depth, thermal, IMU) via a shared embedding space—without explicit paired data for all combinations—yields emergent zero-shot fusion capabilities. This breakthrough relies on contrastive learning that aligns modalities through a common anchor (e.g., image-audio pairs) and transfers alignment to unpaired modalities via transitive inference. Such "emergent binding" reduces the need for exhaustive paired datasets, a long-standing bottleneck in multimodal learning.

2.2 Uncertainty-aware and evidential fusion Classical fusion assumes known confidence, but real-world sensors degrade unpredictably. Recent work onevidential deep learning(Sensoy et al., 2018; extended by Ulmer et al., 2023) treats each modality as a source of subjective logic evidence, producing Dirichlet distributions over class probabilities. Fusion then becomes a Dempster-Shafer combination rule embedded in a differentiable layer. This allows the model to explicitly represent "conflict" between modalities—e.g., when a camera sees a clear road but LiDAR reports an obstacle—and to defer decision-making when evidence is contradictory. In autonomous driving,multimodal evidential fusion(Yang et al., 2024) reduced fatal misclassification rates by 41% compared to softmax-based fusion under sensor noise.

2.3 Graph-based fusion for irregular data Many fusion problems involve non-Euclidean data: social networks, molecular structures, or traffic grids. Graph neural networks (GNNs) now serve as flexible fusion backbones. A notable advance isHeterogeneous Graph Transformer for Fusion(Hu et al., 2023), which models different modalities as node types and learns edge weights that reflect cross-modal dependencies. Applied to urban air-quality monitoring, this method fused satellite aerosol data, meteorological time series, and traffic flow graphs, achieving a 23% improvement in fine-grained PM2.5 prediction over previous convolutional LSTMs. Furthermore,dynamic graph fusion(Zhang et al., 2025) adapts graph topology in real time, dropping unreliable sensor nodes and adding new ones—critical for drone swarms where communication links are transient.

2.4 Generative fusion and synthetic data augmentation A paradigm shift is the use of generative models—especially diffusion models—forfusion-driven synthesis. Instead of only combining existing data, modern systems generate missing modalities from available ones. For example,Diffusion-based Cross-Modal Completion(Saharia et al., 2023) uses a conditional diffusion model to generate high-resolution satellite imagery from sparse radar data, then fuses the synthetic image with real data for flood mapping. This "fusion-before-generation" approach improves downstream segmentation accuracy by 18% in low-visibility conditions. More radically,UniFusion(Li et al., 2024) trains a single diffusion model to generate any missing modality (text, audio, thermal) given any subset of available modalities, effectively turning data fusion into a conditional generation problem. This has profound implications for data imputation in electronic health records, where missing lab tests are common.

3. Technical breakthroughs enabling scalable fusion

3.1 Federated fusion for privacy-preserving systems When data cannot leave local devices (e.g., hospital records, smartphone biometrics), federated learning provides a distributed fusion framework. Recent work byFedFusion(Chen et al., 2024) introduces a server-side "fusion head" that aggregates latent representations from clients without sharing raw data. Crucially, this method usespersonalized normalizationto account for non-IID (non-identically distributed) data across clients—a known failure mode of naive federated averaging. In a multi-hospital study, federated fusion of MRI, genomic, and clinical text data improved tumor classification AUC by 0.07 while maintaining zero data leakage. Complementary work onsecure multiparty computation fusion(Bai et al., 2025) demonstrates that homomorphic encryption can be applied to transformer attention matrices, enabling fusion of proprietary financial datasets across competing institutions.

3.2 Physics-informed fusion For scientific applications, pure data-driven fusion often violates physical laws (e.g., conservation of mass).Physics-informed neural networks(PINNs) have been extended to fusion tasks viamulti-physics constrained loss functions. A landmark study (Karniadakis et al., 2023) fused sparse sensor measurements with a partial differential equation (PDE) model of ocean circulation. The fusion network not only interpolated missing velocities but also enforced the Navier-Stokes equations as a soft constraint, reducing reconstruction error by 3× over purely data-driven methods. Similarly, in structural health monitoring,physics-informed fusion of accelerometer and strain gauge data(Wang et al., 2024) predicts crack propagation with 89% accuracy, using finite-element models as a regularizer.

3.3 Self-supervised and foundation-model fusion The rise of large pre-trained models has enabledzero-shot fusionwithout task-specific training. The key idea is to use a frozen multimodal foundation model (e.g., CLIP, GPT-4V) as a "fusion prior." For instance,CLIP-Fusion(Radford et al., 2023) demonstrated that aligning image and text embeddings can be repurposed to fuse unseen sensor modalities by projecting their embeddings into the CLIP space via a lightweight adapter. This eliminates the need for large paired fusion datasets. More recently,Any-to-Any Foundation Models(e.g., Meta’sCM3leonand Google’sGemini) natively accept interleaved text, image, audio, and video as input tokens, performing fusion in a unified autoregressive framework. These models exhibit emergent reasoning across modalities—e.g., answering "Is the sound of the crowd consistent with the scoreboard image?"—which was previously impossible.

4. Future outlook and open challenges

4.1 Dynamic and lifelong fusion Current fusion models are static after training. Real-world systems requirelifelong fusion—adapting to new sensor types, new data distributions, and new tasks without catastrophic forgetting. Early work oncontinual multimodal learning(Zhai et al., 2024) uses a combination of replay buffers and elastic weight consolidation to retain previously learned fusion patterns. However, scalable solutions for large foundation models remain elusive. We foresee the emergence ofmodular fusion routersthat dynamically select which modality-specific encoders to activate based on input availability and task context.

4.2 Causal fusion Most fusion methods learn correlations, not causal relationships. For high-stakes decisions (e.g., medical diagnosis), we need to knowwhya fusion of symptoms and imaging leads to a conclusion. Recent work oncausal representation learning for multimodal data(Liang et al., 2025) uses structural causal models to disentangle shared causes (e.g., disease) from modality-specific noise (e.g., MRI artifacts). This allows counterfactual reasoning: "If the X-ray had been clean, would the diagnosis change?" We anticipate that causal fusion will become mandatory for regulatory approval in medical and autonomous systems.

4.3 Ethical and governance challenges Fusion amplifies the risk ofinference leakage—combining innocuous data sources (e.g., social media posts + purchase history) to infer sensitive attributes (e.g., health status). The European AI Act now classifies certain fusion applications as "high-risk," requiring explainability and human oversight. Future research must developfused-model interpretabilitytools (e.g.,