The survey, authored by Davi Fileti and André Carvalho of the University of São Paulo, Márcio Basgalupp of the Federal University of São Paulo, and João Gama of the University of Porto, systematically reviews 68 state-of-the-art multimodal data augmentation approaches. Their analysis, published as an open-access article that carries a permanent DOI, examines each method’s underlying methodology, its intended applications, and the predictive performance gains it delivers, while also cataloguing the field’s most pressing open problems, including scalability and the lack of standardized evaluation metrics. The work was supported by Brazil’s CAPES and CNPq and by FAPESP, and the authors report no conflicts of interest.

Multimodal data augmentation, often abbreviated MMDA, is the multimodal extension of a technique that has been a quiet workhorse of machine learning for years. In classical unimodal settings, augmentation means artificially expanding a training dataset by creating modified versions of existing samples—flipping and cropping images, injecting noise into audio clips, or replacing words with synonyms in text. The idea is simple: if the model sees more variations of the underlying signal, it learns features that generalize better and is less likely to memorize quirks of the training set. When several modalities are involved, however, augmentation becomes substantially more intricate. A single training example may consist of an image, its caption, and perhaps an audio track, all of which must remain semantically coherent after transformation. Augment only the image and the caption may no longer describe it; augment both independently and the pairing may drift apart, teaching the model false correspondences. It is precisely this interplay that the new survey dissects.

According to the authors, MMDA addresses three critical challenges that recur across multimodal learning. The first is data scarcity: annotated multimodal datasets are expensive to produce because each sample requires synchronized, aligned information from multiple sources, and for many domain-specific tasks—rare diseases, low-resource languages, industrial fault detection—such data simply does not exist in useful quantities. The second is modality imbalance, a phenomenon in which one modality dominates training data while others are scarce or noisy; models trained under such conditions tend to over-rely on the dominant stream and ignore complementary information that could improve robustness. The third is cross-modal alignment, the technical problem of ensuring that representations from different modalities correspond correctly to one another in a shared semantic space, so that the model understands which caption goes with which image and which transcript accompanies which audio segment. Augmentation strategies, the survey shows, can be designed to target each of these bottlenecks individually or in combination.

The centerpiece of the paper is its proposed taxonomy, a hierarchical classification that organizes the 68 reviewed methods into a unified framework. Rather than treating augmentation techniques as an undifferentiated collection of tricks, the taxonomy groups them according to the nature of the transformations they apply and the role they play in the learning pipeline. Some methods manipulate existing samples—geometric or photometric transformations of images paired with corresponding edits to associated text, for instance—while others generate entirely new synthetic samples, drawing on the growing power of generative models. Still others operate at the level of training strategy, modulating how different modalities are sampled or weighted during optimization rather than transforming the data itself. By placing each published approach within this structure, the authors give researchers a common vocabulary for comparing methods, understanding their strengths and limitations, and identifying which families of techniques are best suited to which problem profiles. This kind of consolidation matters: in fast-moving areas of machine intelligence, surveys that impose order on a chaotic literature often shape the direction of subsequent research as much as any single breakthrough paper.

The performance evidence assembled in the survey makes a compelling case that augmentation is not merely a convenience but frequently a decisive factor in multimodal system quality. The authors document consistent predictive gains across application domains, from vision-language tasks to settings involving audio, sensor streams and structured data, with the largest benefits appearing precisely where data is scarce or imbalanced—the conditions under which multimodal systems are hardest to build in the first place. The survey also emphasizes, however, that gains are not uniform. Methods that excel in one modality combination may transfer poorly to another, and the field currently lacks standardized benchmarks and evaluation metrics, making it difficult to compare results across papers or to know when a reported improvement reflects a genuine methodological advance rather than favorable experimental conditions. Scalability emerges as a parallel concern: several of the most effective augmentation strategies rely on computationally expensive generative components, raising questions about their practicality at industrial scale.

Perhaps the most forward-looking section of the survey concerns the emerging trends it identifies, chief among them the integration of large language models and diffusion processes into the augmentation pipeline. Large language models, with their capacity for fluent text generation and semantic reasoning, are increasingly being used to paraphrase, expand or synthesize textual modalities, and in vision-language contexts to generate descriptions that remain faithful to paired imagery. Diffusion models—generative architectures that learn to create images, audio and other data by gradually denoising random inputs—have opened the possibility of synthesizing photorealistic new multimodal samples on demand, effectively manufacturing training data where none existed. The survey suggests that these generative approaches represent a step change in what augmentation can achieve, moving the field from perturbing existing samples toward open-ended sample synthesis, while also introducing new risks: synthetic data can encode biases present in the generative model, and poorly calibrated generation can pollute training sets with subtly incoherent multimodal pairs.

For practitioners, the taxonomy serves a practical triage function. A team facing a scarcity of labeled paired data can consult the survey to identify which augmentation families have demonstrated success in analogous situations, whether through cross-modal translation, generative synthesis or strategy-level interventions such as modality dropout, in which a model is deliberately trained with certain inputs withheld so that it does not become dependent on any single stream. A team wrestling with modality imbalance can find methods designed to rebalance representation learning across streams. The survey’s analysis of each paper’s methodology and reported performance provides an entry point into the primary literature that is dramatically more efficient than keyword searching a field that has fragmented across computer vision, natural language processing, speech and multimodal fusion communities.

The authors are candid about the limitations of the current state of the art and the research directions that follow from them. Beyond the need for standardized evaluation protocols and scalable implementations, the survey points toward a deeper conceptual question: what does it mean for an augmented multimodal sample to be valid? In unimodal augmentation, plausibility is often sufficient—a flipped photograph of a cat is still a cat. In multimodal data, validity is a relational property, requiring that all transformed components remain mutually consistent. Formalizing and measuring that consistency, the authors suggest, is one of the fundamental open problems the field must solve. They also anticipate that the boundary between augmentation and generation will continue to blur as foundation models and diffusion-based systems become standard components of training pipelines, potentially redefining what constitutes a training dataset at all.

The significance of the work extends beyond the machine learning research community. Multimodal systems are now embedded in consequential applications—medical imaging and clinical decision support, autonomous driving, content moderation, accessibility technology for users who communicate across text, speech and vision—and the reliability of these systems depends directly on how well their models handle data heterogeneity. By cataloguing the techniques that make multimodal training feasible under real-world data constraints, the survey addresses one of the quiet but essential enabling layers of modern artificial intelligence. The paper, received in February 2026 and accepted in July before its August publication, is already available in a citable early-access form, reflecting the journal’s practice of sharing peer-reviewed accepted research quickly. For a field that has produced dozens of novel augmentation methods in the past few years without a common framework for understanding them, the arrival of a systematic, taxonomy-driven review marks a moment of consolidation—and, if its identified trends are any guide, a launching point for the next phase of multimodal learning research, one in which large language models and diffusion processes are not just tools within systems but engines of the data those systems learn from.