Multi-source Multimodal Progressive Domain Adaptation

• MMPDA is a framework that integrates multiple labeled source domains and various data modalities using progressive, curriculum-based alignment.
• It employs multimodal representation alignment, source aggregation, and staged adaptation techniques to transfer audio-visual, textual, or behavioral features effectively.
• Instantiations in audio-visual deception detection and CLIP-based unsupervised adaptation demonstrate improved accuracy and robustness against domain shifts.

Multi-source Multimodal Progressive Domain Adaptation (MMPDA) denotes a class of domain adaptation methods that transfer knowledge from multiple labeled source domains to a target domain by jointly exploiting more than one modality and by organizing adaptation as a staged or curriculum-like process rather than a one-shot alignment. In the supplied literature, the term is used explicitly for audio-visual deception detection, where a framework “transfers the audio-visual knowledge from diverse source domains to the target domain” and “gradually align[s] source and the target domain at both feature and decision levels” (Lin et al., 18 Aug 2025). The same conceptual triad—multi-source, multimodal, and progressive—also appears in CLIP-based unsupervised adaptation, where MP$^2$A is described as “exactly the type of method” instantiated by MMPDA, and in progressive multi-view segmentation, which treats progressive adaptation over increasingly complex views as a template that can be extended to multi-source and multimodal settings (Chen et al., 31 Jul 2025, Nie et al., 5 Feb 2026).

1. Definition and conceptual scope

MMPDA is best understood as the intersection of three established dimensions of adaptation. The multi-source dimension assumes multiple labeled source domains and one target domain rather than a single source; the multimodal dimension assumes that adaptation uses multiple modalities or a joint multimodal embedding space rather than a single unimodal feature stream; the progressive dimension assumes that alignment or target adaptation is performed incrementally, for example from easy to hard samples, from low-perturbation to high-perturbation views, or through staged increases in adaptation strength (Chen et al., 31 Jul 2025, Zhang et al., 24 Jun 2025, Lin et al., 18 Aug 2025).

Within the supplied material, this concept has both a narrow and a broad sense. In the narrow sense, MMPDA refers to the specific framework for audio-visual deception detection that uses audio, visual frames, and behavioral features, fuses them with MulT, and combines CORAL, MDD, entropy regularization, and adversarial learning under a progressive schedule (Lin et al., 18 Aug 2025). In the broader sense, MMPDA is a design pattern abstracted from multiple lines of work: CLIP-based multi-source unsupervised adaptation with progressive curriculum over pseudo-labeled target classes (Chen et al., 31 Jul 2025), progressive multi-view adaptation for cross-domain few-shot segmentation (Nie et al., 5 Feb 2026), and multi-modality domain adaptation with progressive modality cooperation and privileged-modality generation (Zhang et al., 24 Jun 2025).

This broader reading is supported explicitly in the source material. MP$^2$A states that “Multi-source Multimodal Progressive Domain Adaptation (MMPDA) is exactly the type of method MP$^2$A instantiates,” because it is multi-source, multimodal through CLIP image–text modeling, and progressive through a “learn, refine, rehearse” curriculum (Chen et al., 31 Jul 2025). The segmentation paper similarly states that its “Multi-view Progressive Adaptation” can inspire an MMPDA framework by treating different sources or modalities as views and applying progressive augmentation and dual-chain prediction across them (Nie et al., 5 Feb 2026).

2. Formal problem settings

A recurring formal setting for MMPDA is multi-source unsupervised domain adaptation. In MP$^2$A, there are $N-1$ labeled source domains and one unlabeled target domain, all sharing the same set of $K$ semantic classes. The sources are

$$\{\mathcal{D}_s^i\}_{i=1}^{N-1}, \qquad \mathcal{D}_s^i = \{(x_s^{i,n}, y_s^{i,n})\}_{n=1}^{|\mathcal{D}_s^i|}$$

and the target is

$$\mathcal{D}_t = \{x_t^{m}\}_{m=1}^{|\mathcal{D}_t|},$$

with no labels, only pseudo-labels $\hat{y}_t^m$ generated during training (Chen et al., 31 Jul 2025). The objective is to predict correct labels for target samples by transferring knowledge from all sources under domain shift.

When the multimodal backbone is CLIP, the multimodal aspect is formalized through an image encoder $\phi(\cdot)$ and text encoder $^2$0, both mapping into a shared embedding space. Zero-shot classification is written as

$^2$1

and MP$^2$2A extends this by using domain-aware prompts and parameter-efficient fine-tuning of the visual encoder (Chen et al., 31 Jul 2025). This yields a concrete multimodal MS-UDA formulation in which text prompts, visual prompts, LoRA, or adapters are part of the adaptation mechanism.

A different but complementary formalization appears in audio-visual deception detection. There, each sample contains acoustic input $^2$3, visual input $^2$4, and a binary label $^2$5. Source domains are labeled,

$^2$6

while the target domain is unlabeled,

$^2$7

The goal is to transfer “audio-visual knowledge from multiple source domains to an unlabeled target domain,” while bridging shifts in speech, appearance, and behavior distributions (Lin et al., 18 Aug 2025).

The multi-modality domain adaptation literature also provides a privileged-information variant. PMC defines MMDA when both source and target have all modalities, and MMDA-PI when some target modalities are missing, so that source-only modalities must be generated for the target with a dedicated network (Zhang et al., 24 Jun 2025). This suggests that MMPDA need not require identical modality availability across all domains; a plausible implication is that modality completion can be treated as part of the adaptation process when target observations are incomplete.

3. Core methodological patterns

Three mechanisms recur across the supplied work and collectively characterize MMPDA: multimodal representation alignment, multi-source aggregation, and progressive target adaptation.

The multimodal representation component is realized in different ways. In CLIP-based methods, multimodality is image–text: learnable prompts and visual PEFT modules are trained in a shared image–text embedding space, and multiple prompt sets $^2$8 are aligned by an autoencoder and a cross-prompt consistency loss $^2$9 (Chen et al., 31 Jul 2025). In audio-visual deception detection, multimodality is audio, visual frames, and behavioral features; modality-specific encoders produce $^2$0, $^2$1, and $^2$2, which are fused by MulT into a shared multimodal representation

$^2$3

followed by an MLP classifier (Lin et al., 18 Aug 2025). In PMC, multimodality is handled by separate modality-specific feature extractors, classifiers, and domain discriminators, together with modality-specific and modality-integrated pseudo-label selection (Zhang et al., 24 Jun 2025).

The multi-source component can take the form of explicit source-specific modules or collaborative source aggregation. MP$^2$4A trains a separate CLIP-based model $^2$5 for each source domain and ensembles their outputs to generate target pseudo-labels by average-confidence voting or majority voting (Chen et al., 31 Jul 2025). ML-MSDA instead constructs $^2$6 branch networks, one per source, plus one guidance network trained on the union of all sources, and uses a symmetric Jensen–Shannon style mutual learning loss

$^2$7

to regularize target predictions across branches and guidance (Li et al., 2020). LanDA uses domain-specific augmenters $^2$8, multimodal Wasserstein alignment across extended source domains, and OT-based source weights computed from text-domain descriptions rather than target images (Wang et al., 2024).

The progressive component is the defining departure from one-shot adaptation. In MP$^2$9A, progression is organized over class clusters ordered by a difficulty score

$^2$0

with training following an easy-to-hard sequence and a “learn phase,” “refine phase,” and “rehearse phase” (Chen et al., 31 Jul 2025). In PMC, progression is driven by self-paced pseudo-label selection ratios $^2$1 and $^2$2, so that only reliable target samples are used early and harder samples are included later (Zhang et al., 24 Jun 2025). In audio-visual MMPDA, progression is expressed through domain-by-domain feeding, the joint use of feature-level and decision-level alignment losses, and a gradient reversal schedule whose strength increases with training progress $^2$3 (Lin et al., 18 Aug 2025). In MPA for segmentation, progression is created by cumulative strong augmentations and by increasing the number of views as performance saturates, which is explicitly framed as a progressive domain adaptation strategy (Nie et al., 5 Feb 2026).

A concise way to compare representative mechanisms is the following.

Method	Setting	Progressive mechanism
MP$^2$4A (Chen et al., 31 Jul 2025)	CLIP-based MS-UDA	Easy-to-hard class clusters; learn–refine–rehearse
PMC (Zhang et al., 24 Jun 2025)	MMDA / MMDA-PI	Self-paced MSS and MIS pseudo-label selection
MMPDA for deception (Lin et al., 18 Aug 2025)	Audio-visual MSDA	GRL schedule, domain-by-domain training, staged adaptation

These patterns suggest that MMPDA is less a single architecture than a compositional recipe: use multiple sources without collapsing them, exploit multiple modalities in a shared or cooperating representation space, and avoid injecting all uncertain target information at once.

4. Representative instantiations

The most explicit instantiation is the framework titled “Multi-source Multimodal Progressive Domain Adaption for Audio-Visual Deception Detection” (Lin et al., 18 Aug 2025). It uses ResNet18+LSTM encoders for audio and visual streams, an additional encoder for behavioral features extracted from Emonet and OpenFace, and MulT for cross-modal fusion. The adaptation loss is

$^2$5

combined with multimodal and unimodal task losses in

$^2$6

On the MMDD Challenge, the multi-source setting DOLOS + MDPE $^2$7 MMDD reaches 60.43% accuracy and 56.99% F1-score, and the paper reports Top-2 placement, with gains over baseline of +5.51 in accuracy and +7.86 in F1 (Lin et al., 18 Aug 2025). Those numbers are specific to the competition setting and establish the term MMPDA in a concrete multimodal benchmark.

A second representative line is CLIP-based multi-source unsupervised adaptation. MP$^2$8A defines MMPDA conceptually and instantiates it through source-specific prompt sets, domain-specific and class-specific textual tokens, shared visual PEFT, and an easy-to-hard curriculum over target class clusters (Chen et al., 31 Jul 2025). Its prompt alignment stage optimizes

$^2$9

where $N-1$0 reconstructs prompt sets through an autoencoder and $N-1$1 enforces cross-prompt consistency. The method is evaluated on ImageCLEF, Office-Home, and DomainNet, with approximate average accuracies of 94.3% on ImageCLEF, 91.8% on Office-Home, and 64.1% on DomainNet for the best PEFT variants (Chen et al., 31 Jul 2025). Because the paper explicitly states that MP$N-1$2A is “exactly the type of method” MMPDA refers to, it supplies an image–text realization of the concept.

A third representative line comes from progressive multimodal cooperation. PMC is not multi-source, but it provides a detailed progressive multimodal adaptation engine. Under MMDA, it adds modality-specific sample selection and modality-integrated sample selection to a DANN-style backbone, producing a target pseudo-label loss

$N-1$3

that combines per-modality confident samples with cross-modality consensus samples (Zhang et al., 24 Jun 2025). Under MMDA-PI, PMC-PI supplements this with an MMG network that generates missing target modalities by minimizing a reconstruction loss $N-1$4 and an adversarial loss $N-1$5 (Zhang et al., 24 Jun 2025). The paper’s own framing makes this relevant to MMPDA because it demonstrates how progressive pseudo-labeling, modality cooperation, and missing-modality synthesis can be combined in a unified adaptation loop.

A fourth representative direction is “Cross-Domain Few-Shot Segmentation via Multi-view Progressive Adaptation” (Nie et al., 5 Feb 2026). It is not multi-source in the strict formulation, but it formalizes progression through Hybrid Progressive Augmentation and Dual-chain Multi-view Prediction. The method states that it “progressively adapts few-shot capability to target domains from both data and strategy perspectives,” and explicitly proposes that its design can be extended so that “different sources or modalities become ‘views’” in a multi-source multimodal progressive framework (Nie et al., 5 Feb 2026). This is not an implemented MMPDA method, but it is a direct conceptual bridge from progressive multi-view adaptation to MMPDA.

5. Relation to adjacent research areas

MMPDA sits at the intersection of several adjacent literatures, and the supplied papers clarify where it departs from each.

Relative to classical multi-source domain adaptation, MMPDA inherits the central challenge that pooling all sources can be harmful. The surveys on multi-source domain adaptation emphasize that naïve “source-combining” often underperforms because source–source interference exacerbates mismatch, and they identify source weighting, domain-specific branches, latent-space transformation, and intermediate-domain generation as core strategies (Zhao et al., 2020, Zhao et al., 2024). ML-MSDA exemplifies one response by pairing each source with the target in a separate conditional adversarial branch and aligning all branches with a guidance network through mutual learning (Li et al., 2020). T-SVDNet exemplifies another by stacking prototypical similarity matrices from all domains into a tensor and imposing a tensor-low-rank constraint so that “high-order correlations among multiple domains and categories are fully explored” (Li et al., 2021). These methods are multi-source but unimodal and non-progressive.

Relative to multimodal domain adaptation, MMPDA adds either multiple sources or a progressive schedule. PMC is explicitly “a new generic multi-modality domain adaptation framework,” but it assumes one source and one target (Zhang et al., 24 Jun 2025). LanDA is multi-source and multimodal because it uses a frozen visual-language foundation model, text descriptions of source and target domains, and multimodal optimal transport, yet it is not explicitly progressive beyond a two-stage training process (Wang et al., 2024). MS-DePro is multi-source and multimodal for object detection, using RGB, depth, and text, but its “progressive” behavior is largely implicit in its burn-in and teacher–student schedule rather than being formulated as a target curriculum (Lee et al., 13 May 2026).

Relative to progressive domain adaptation, MMPDA adds multi-source and multimodal structure. Progressive Domain Augmentation constructs virtual intermediate domains by interpolating source and target features and then performs multiple subspace alignment on the Grassmann manifold, which is a purely progressive but unimodal and single-source mechanism (Hua et al., 2020). MPA similarly implements a strong progressive schedule over augmented views and supervision paths, but in a few-shot segmentation context (Nie et al., 5 Feb 2026). These works show that progression can be data-centric or curriculum-centric even when multi-source or multimodal factors are absent.

A useful summary is that MMPDA combines the branch-aware source handling of MSDA, the representation or prompt cooperation of multimodal DA, and the sample- or curriculum-aware scheduling of progressive DA. The supplied literature does not present a single unified theory covering all three simultaneously. This suggests that MMPDA is currently better described as an overview of mechanisms than as a settled formal subfield.

6. Limitations, open questions, and likely directions

Several limitations recur across the supplied work and define the current boundary of MMPDA research. One limitation is that many methods are only partially triadic: PMC is multimodal and progressive but not multi-source (Zhang et al., 24 Jun 2025); LanDA is multi-source and multimodal but only implicitly progressive (Wang et al., 2024); audio-visual MMPDA is multi-source, multimodal, and progressive, but in a task-specific binary deception setting (Lin et al., 18 Aug 2025). This suggests that a general-purpose MMPDA framework remains an open synthesis problem.

A second limitation concerns static curricula and heuristic confidence rules. MP$N-1$6A notes that its curriculum is static because it is determined before training using initial CLIP embeddings, and it identifies heuristic thresholds $N-1$7 and $N-1$8 as potential weaknesses (Chen et al., 31 Jul 2025). PMC likewise relies on self-paced percentages $N-1$9 and $K$0 updated from source accuracy trends rather than from a principled posterior uncertainty model (Zhang et al., 24 Jun 2025). A plausible implication is that future MMPDA systems may combine progressive schedules with stronger uncertainty estimation or calibration mechanisms, although such a design would go beyond the explicit implementations in the supplied papers.

A third limitation is scalability and optimization complexity when many sources, modalities, or classes are combined. T-SVDNet shows that modeling high-order inter-domain category structure can be effective, but it requires tensor construction, FFT-based T-SVD, and alternating optimization (Li et al., 2021). Graphical approaches such as CRF-MSDA and MRF-MSDA also scale with prototype graphs and domain-category combinations (Xu et al., 2021). This suggests that extending MMPDA to large multimodal settings may require more efficient tensor or graph updates, especially when domains arrive sequentially.

A fourth limitation is uneven treatment of modalities. The audio-visual deception framework explicitly notes that it aligns only fused multimodal representations and does not perform explicit unimodal-domain alignment, so residual mismatches at the unimodal level may weaken the fused representation (Lin et al., 18 Aug 2025). MS-DePro similarly treats RGB as domain-specific and depth/text as domain-agnostic, which is powerful for the studied detection setting but presupposes that those modalities have the stated invariance properties (Lee et al., 13 May 2026). This suggests that MMPDA is likely to remain task- and modality-dependent until broader comparative evidence accumulates.

Finally, the surveys identify multi-modal MDA and incremental or online MDA as open directions, and these two directions together closely match the natural next stage of MMPDA (Zhao et al., 2020, Zhao et al., 2024). The literature already supplies most ingredients separately: source weighting and branch-guidance architectures (Li et al., 2020), multimodal OT and text-based domain weighting (Wang et al., 2024), progressive pseudo-label curricula (Zhang et al., 24 Jun 2025), and explicit progressive alignment schedules (Chen et al., 31 Jul 2025, Nie et al., 5 Feb 2026). The main unresolved question is how to combine them without destabilizing training or amplifying negative transfer.