MMLNet: Multi-Expert Incomplete Modality Learning

• The paper introduces a multi-expert framework that fuses text, image, and cross-modal predictions to robustly detect fake news under incomplete modality conditions.
• It employs lightweight Incomplete Modality Adapters and a label-aware contrastive loss to stabilize feature representations when modalities are partially missing.
• Empirical results show that MMLNet maintains high accuracy and minimal performance degradation compared to state-of-the-art methods even under heavy modality missingness.

Searching arXiv for the specific topic and closely related incomplete-modality multimodal learning papers. arXiv search query: "Multi-Expert Modality-Incomplete Learning Network MMLNet incomplete modality multimodal" Multi-Expert Modality-Incomplete Learning Network (MMLNet) denotes an expert-based framework for multimodal learning under missing-modality conditions. In the explicitly titled usage, MMLNet is introduced as a novel generic and robust multimodal fusion strategy for multimodal fake news detection with incomplete modality, combining Multi-Expert Collaborative Reasoning, Incomplete Modality Adapters, and Modality Missing Learning with a label-aware adaptive weighting strategy (Zhou et al., 7 Oct 2025). The same label also appears in a self-contained description attached to work on residual-guided expert specialization, where MMLNet is presented as a sparse Mixture-of-Experts backbone for incomplete multimodal learning that closely follows the ideas of MARS (Baek et al., 29 Jun 2026). The term therefore refers to a family of expert-routed responses to modality incompleteness rather than to a single universally fixed architecture.

1. Definition, scope, and naming

In multimodal fake news detection, MMLNet is defined around the observation that real-world social media posts often lose words or image patches as they propagate, while most existing multimodal fake-news detectors assume that both modalities are always fully present. The framework is designed to explicitly train on and test with incomplete text/image pairs, dynamically leverage multiple experts to compensate for missing signals, introduce lightweight adapters to reshape feature distributions when modalities are partially observed, and employ a label-aware contrastive loss to tighten feature clusters of same-label samples while pushing apart opposite-label ones in the presence of missing data (Zhou et al., 7 Oct 2025).

The terminology is not fully unique across the literature. In the residual-specialization line, the same name is used for a model built on a sparse Mixture-of-Experts backbone, where each training sample is presented in complete and incomplete forms, a residual $r = z^{\rm full} - z^{\rm inc}$ is computed, and expert routing is guided by this missingness-induced representational shift (Baek et al., 29 Jun 2026). A common misconception is that all instances of “MMLNet” denote the same model family. The published material instead indicates two distinct usages: one centered on multimodal fake news detection with experts, adapters, and contrastive learning, and another centered on residual-guided expert specialization.

2. Multi-Expert Collaborative Reasoning in multimodal fake news detection

The fake-news MMLNet maintains three parallel expert modules: $f_h$ for the text $[CLS]$ token, $f_r$ for the image $[CLS]$ token, and $f_f$ for the cross-modal transformer that ingests both text and image tokens. With $H = \mathrm{CLIP\_text}(x^T)$ and $R = \mathrm{CLIP\_image}(x^V)$, the unimodal features are extracted as

$$F_T = f_A^T(h_{[CLS]}), \qquad F_I = f_A^I(r_{[CLS]}).$$

Each expert then produces a logits distribution,

$$y^h = \mathrm{softmax}(W_T F_T + b_T), \qquad y^r = \mathrm{softmax}(W_I F_I + b_I),$$

while the fused expert uses

$f_h$0

where $f_h$1 and $f_h$2 result from a self-attention block over $f_h$3 (Zhou et al., 7 Oct 2025).

The three experts are combined by learned dynamic routing: $f_h$4 In generic notation,

$f_h$5

with $f_h$6, $f_h$7, and $f_h$8. This formulation places MMLNet within expert-routing approaches in which compensation for missing information arises from adaptive combination of complementary expert predictions rather than from explicit reconstruction.

A central interpretive point is stated directly in the model description: unlike methods that learn a single joint representation or attempt to reconstruct missing modalities, MMLNet models three complementary experts and fuses them via learned gating, and it does not try to “hallucinate” missing pixels or words. This suggests a design preference for robustness through expert diversity and missing-aware adaptation of the feature space rather than through generative completion (Zhou et al., 7 Oct 2025).

3. Incomplete Modality Adapters and label-aware contrastive learning

A second core component is the use of Incomplete Modality Adapters. Motivated by the observation that large pre-trained encoders may over-fit when data are missing, MMLNet attaches lightweight residual adapters $f_h$9 and $[CLS]$0 to the unimodal $[CLS]$1 features: $[CLS]$2 The final adapted features are

$[CLS]$3

with $[CLS]$4, $[CLS]$5, and $[CLS]$6 a small residual ratio (Zhou et al., 7 Oct 2025). The explicit purpose is to learn a “missing-aware” correction without overwriting the original pre-trained embedding.

MMLNet further augments classification with a contrastive objective termed Modality Missing Learning. For an anchor sample $[CLS]$7, the positive and negative sets are defined as

$[CLS]$8

Let $[CLS]$9 be a small MLP projection and $f_r$0 a temperature. The vanilla multimodal contrastive loss is

$f_r$1

Because not all positives and negatives are equally informative when modalities are missing, the model re-weights each pair by cosine similarity: $f_r$2 The resulting label-aware contrastive loss becomes

$f_r$3

This loss is applied in parallel to the three expert representations $f_r$4 (Zhou et al., 7 Oct 2025).

Within the stated design, the adapters and the label-aware contrastive objective serve different but complementary roles. The adapters compensate for the new feature distribution caused by missing modalities, whereas the contrastive term is used to stabilize representation geometry under partial observability. A plausible implication is that MMLNet separates distributional correction from representation discrimination rather than treating both as a single fusion problem.

4. Optimization, training protocol, and reported performance

For each expert $f_r$5, the classification loss is

$f_r$6

The total loss is

$f_r$7

where $f_r$8 and $f_r$9 balance classification and contrastive terms (Zhou et al., 7 Oct 2025).

The training steps are specified as follows: precompute random masks for text and image drops; for each batch, apply missing masks to input text tokens and image patches; encode via CLIP_text and CLIP_image; pass $[CLS]$0 tokens through adapters to get $[CLS]$1 and $[CLS]$2; run a cross-modal transformer to get $[CLS]$3 and $[CLS]$4; compute $[CLS]$5, $[CLS]$6, $[CLS]$7 and fuse via $[CLS]$8 to obtain $[CLS]$9; compute $f_f$0 and sample $f_f$1 to compute weighted $f_f$2; then back-propagate $f_f$3 and update parameters, with optional lower-rate CLIP fine-tuning (Zhou et al., 7 Oct 2025). The typical hyperparameters used on all three datasets are AdamW, $f_f$4 with encoder $f_f$5, weight-decay $f_f$6, batch size $f_f$7, epochs $f_f$8, dropout $f_f$9, $H = \mathrm{CLIP\_text}(x^T)$0, $H = \mathrm{CLIP\_text}(x^T)$1, and loss weights $H = \mathrm{CLIP\_text}(x^T)$2, $H = \mathrm{CLIP\_text}(x^T)$3.

The reported datasets are Weibo, Weibo21, and Pheme, with modality-missing scenarios defined by text/image missing rates $H = \mathrm{CLIP\_text}(x^T)$4 and $H = \mathrm{CLIP\_text}(x^T)$5. Evaluation uses Accuracy (ACC), macro-F1, and AUC, and baselines include NSLM, MIMoE, Qwen-VL-7B, and LLaVA-13B (Zhou et al., 7 Oct 2025). Under complete modalities, the results are reported as follows: Weibo baseline best ACC $H = \mathrm{CLIP\_text}(x^T)$6 with MMLNet ACC $H = \mathrm{CLIP\_text}(x^T)$7; Weibo21 baseline best ACC $H = \mathrm{CLIP\_text}(x^T)$8 with MMLNet ACC $H = \mathrm{CLIP\_text}(x^T)$9; Pheme baseline best ACC $R = \mathrm{CLIP\_image}(x^V)$0 with MMLNet ACC $R = \mathrm{CLIP\_image}(x^V)$1. Under heavy missingness, exemplified by text down $R = \mathrm{CLIP\_image}(x^V)$2 and image down $R = \mathrm{CLIP\_image}(x^V)$3, MMLNet degrades by $R = \mathrm{CLIP\_image}(x^V)$4 on Weibo versus $R = \mathrm{CLIP\_image}(x^V)$5 for MIMoE. The paper characterizes these trends as demonstrating superior performance compared to state-of-the-art methods while maintaining relative simplicity (Zhou et al., 7 Oct 2025).

5. Residual-guided MMLNet and the connection to MARS

A second, distinct formulation uses the name MMLNet for a residual-guided expert-specialization architecture. In that description, each sample $R = \mathrm{CLIP\_image}(x^V)$6 is presented during training in two forms: a complete input $R = \mathrm{CLIP\_image}(x^V)$7 and an incomplete input $R = \mathrm{CLIP\_image}(x^V)$8. These are passed through a shared feature-fusion encoder $R = \mathrm{CLIP\_image}(x^V)$9 with two heads,

$$F_T = f_A^T(h_{[CLS]}), \qquad F_I = f_A^I(r_{[CLS]}).$$0

and the residual

$$F_T = f_A^T(h_{[CLS]}), \qquad F_I = f_A^I(r_{[CLS]}).$$1

is used to capture how the fused representation shifts when modalities go missing (Baek et al., 29 Jun 2026).

The routing system comprises a residual router and a feature router. The residual router produces clean logits

$$F_T = f_A^T(h_{[CLS]}), \qquad F_I = f_A^I(r_{[CLS]}).$$2

a noise scale

$$F_T = f_A^T(h_{[CLS]}), \qquad F_I = f_A^I(r_{[CLS]}).$$3

and noisy logits

$$F_T = f_A^T(h_{[CLS]}), \qquad F_I = f_A^I(r_{[CLS]}).$$4

followed by a top-$$F_T = f_A^T(h_{[CLS]}), \qquad F_I = f_A^I(r_{[CLS]}).$$5 softmax

$$F_T = f_A^T(h_{[CLS]}), \qquad F_I = f_A^I(r_{[CLS]}).$$6

The feature router sees only $$F_T = f_A^T(h_{[CLS]}), \qquad F_I = f_A^I(r_{[CLS]}).$$7,

$$F_T = f_A^T(h_{[CLS]}), \qquad F_I = f_A^I(r_{[CLS]}).$$8

and is trained to imitate the residual router (Baek et al., 29 Jun 2026).

The discrepancy-aware noise regularization uses

$$F_T = f_A^T(h_{[CLS]}), \qquad F_I = f_A^I(r_{[CLS]}).$$9

to measure per-expert discrepancy, sorts experts by decreasing $$y^h = \mathrm{softmax}(W_T F_T + b_T), \qquad y^r = \mathrm{softmax}(W_I F_I + b_I),$$0, and defines

$$y^h = \mathrm{softmax}(W_T F_T + b_T), \qquad y^r = \mathrm{softmax}(W_I F_I + b_I),$$1

The training objective combines classification, routing imitation, load-balancing, and noise regularization: $$y^h = \mathrm{softmax}(W_T F_T + b_T), \qquad y^r = \mathrm{softmax}(W_I F_I + b_I),$$2 At inference, only incomplete inputs are used: $$y^h = \mathrm{softmax}(W_T F_T + b_T), \qquad y^r = \mathrm{softmax}(W_I F_I + b_I),$$3 and prediction is made without the residual router or the full-modality encoder (Baek et al., 29 Jun 2026).

The reported results for this residual-guided MMLNet are average ACER from $$y^h = \mathrm{softmax}(W_T F_T + b_T), \qquad y^r = \mathrm{softmax}(W_I F_I + b_I),$$4 to $$y^h = \mathrm{softmax}(W_T F_T + b_T), \qquad y^r = \mathrm{softmax}(W_I F_I + b_I),$$5 on CASIA-SURF, hardest-setting IR-only error from $$y^h = \mathrm{softmax}(W_T F_T + b_T), \qquad y^r = \mathrm{softmax}(W_I F_I + b_I),$$6 to $$y^h = \mathrm{softmax}(W_T F_T + b_T), \qquad y^r = \mathrm{softmax}(W_I F_I + b_I),$$7, mean IoU from $$y^h = \mathrm{softmax}(W_T F_T + b_T), \qquad y^r = \mathrm{softmax}(W_I F_I + b_I),$$8 to $$y^h = \mathrm{softmax}(W_T F_T + b_T), \qquad y^r = \mathrm{softmax}(W_I F_I + b_I),$$9 on MCubeS, average accuracy from $f_h$00 to $f_h$01 on CREMA-D, and average accuracy from $f_h$02 to $f_h$03 on UPMC Food-101 (Baek et al., 29 Jun 2026). Since the associated paper title is “Residual-Guided Expert Specialization for Incomplete Multimodal Learning,” a straightforward interpretation is that this MMLNet label functions as a descriptive architecture name rather than the principal published title.

6. Position within incomplete-modality expert learning

The broader literature represented here shows several recurring design patterns for modality-incomplete learning: expert partitioning, sparse routing, parameter-efficient adaptation, alignment across modalities, and task- or missingness-aware inference. The following comparison organizes the main frameworks mentioned in the same research area.

Framework	Setting	Central mechanism
DeLo (Liu et al., 2 Mar 2026)	Continual Missing Modality Learning	Dual-decomposed low-rank experts, Cross-Modal Guided Routing, Task-Key Memory
MoSARe (Moradinasab et al., 12 Mar 2025)	Incomplete multimodal healthcare learning	Mixture of Experts, Symmetric Aligning, and Reconstruction
MaMOL (Gao et al., 14 Nov 2025)	Remote sensing modality-missing classification	Missing-aware Mixture-of-Loras with dual routing
MARS / residual-guided MMLNet (Baek et al., 29 Jun 2026)	Incomplete multimodal classification and segmentation	Residual-guided expert specialization with dual routers
MMLNet (Zhou et al., 7 Oct 2025)	Multimodal fake news detection with incomplete modality	Multi-Expert Collaborative Reasoning, Incomplete Modality Adapters, label-aware contrastive learning

DeLo is presented as the first framework to leverage a novel dual-decomposed low-rank expert architecture for CMML, with separate visual and textual factor pools, a Cross-Modal Guided Routing strategy, and a Task-Key Memory for task-agnostic inference (Liu et al., 2 Mar 2026). MoSARe integrates expert selection, cross-modal attention, reconstruction, and contrastive alignment to handle incomplete multimodal healthcare data, including histopathology whole-slide images, RNA-Seq profiles, and clinical text reports (Moradinasab et al., 12 Mar 2025). MaMOL reformulates modality missing as a multi-task learning problem and combines a task-oriented dynamic router with a modality-specific-shared static router in a Missing-aware Mixture-of-Loras design (Gao et al., 14 Nov 2025). MARS, which the residual-guided MMLNet closely follows, uses a privileged residual signal derived from complete and incomplete inputs to guide expert specialization and introduces discrepancy-aware noise regularization to bridge the train-test router gap (Baek et al., 29 Jun 2026).

Placed against these systems, the fake-news MMLNet is comparatively focused in modality scope and application domain: it operates on text-image pairs, retains a three-expert structure rather than a large expert pool, and couples expert collaboration with residual adapters and label-aware contrastive reweighting instead of reconstruction or low-rank adaptation (Zhou et al., 7 Oct 2025). This suggests that “MMLNet” is best understood not as a single canonical architecture, but as an instance of a broader movement toward expert-routed multimodal systems that are explicitly engineered for incomplete observations.