Modal Consistency Dynamic Balance (MCDB)

• MCDB is a multimodal learning mechanism that leverages cross-modal consistency to adjust modality contributions, ensuring that no dominant modality overshadows noisier signals like EEG.
• It uses dynamic gradient modulation with adaptive suppression factors computed from similarity statistics to stabilize and align heterogeneous features during training.
• Empirical results demonstrate that MCDB improves accuracy and robustness in multimodal tasks, with notable gains in EEG–image–text decoding and related fusion challenges.

Modal Consistency Dynamic Balance (MCDB) denotes a multimodal learning mechanism in which cross-modal consistency is used to regulate the relative influence of different modalities during training or fusion. In the supplied literature, the term appears explicitly in EEG–image–text decoding, where MCDB is introduced as a dynamic gradient-balancing strategy that suppresses dominant modalities and stabilizes alignment to low-SNR EEG features (Sun et al., 3 Sep 2025). Closely related mechanisms appear under other names in multimodal large models and cross-modal clustering, where semantic consistency is combined with adaptive weighting or balance regularization at feature or cluster level (Tanaka et al., 15 Jun 2025, Li et al., 2 Aug 2025). A broader thesis on uncertainty-resilient multimodal learning is aligned with the same emphasis on semantic alignment and robustness, although its available skeleton does not expose the actual dynamic-balancing mechanism (Jang, 18 Nov 2025).

1. Conceptual scope and motivating problem

In its explicit multimodal-learning usage, MCDB addresses a recurrent failure mode of multimodal optimization: strong modalities dominate the shared representation, while weaker or noisier modalities are under-optimized. In the HMAVD framework for decoding visual neural representations, the motivating case is EEG–image–text learning. EEG is described as noisy and low-dimensional in semantics, whereas image and especially text features are strong and pre-trained. The reported gradient analysis shows that text gradients persistently rise in later epochs, image gradients plateau, and EEG gradients remain small and noisy. MCDB is introduced to counteract this imbalance by monitoring how strongly image and text features influence EEG alignment, deriving adaptive modality weights, and multiplying those weights into encoder-gradient updates (Sun et al., 3 Sep 2025).

Within that formulation, “modal consistency” does not mean simple feature similarity in the abstract. It refers specifically to consistency of contributions between EEG and the other modalities in the shared space. The intended outcome is that EEG should be aligned to both image and text in a way that neither visual nor linguistic information overwhelms EEG. MCDB operationalizes this through EEG–image and EEG–text similarity statistics, using them to detect dominance and suppress it dynamically.

The broader multimodal literature supplied here suggests that this coupling of semantic consistency and adaptive balancing is not confined to EEG. Dynamic Modality Scheduling in MLLMs evaluates each modality by confidence, uncertainty, and semantic consistency before assigning soft fusion weights, while Self-Enhanced Image Clustering combines image–text consistency with a dynamic balancing regularizer over cluster usage. These are not identical mechanisms, but they instantiate the same design intuition: consistency should guide how modal influence is balanced.

2. Canonical mathematical formulation in HMAVD

The most explicit and fully specified MCDB formulation in the supplied material is the one embedded in HMAVD for EEG-based visual decoding (Sun et al., 3 Sep 2025). For each sample $i$ and modality $m \in \{\text{eeg}, \text{img}, \text{text}\}$, the feature extractor is

$$\mathbf{f}_i^{(m)} = \psi^{(m)}(\mathbf{x}_i^{(m)}; \boldsymbol{\theta}^{(m)}),$$

and for a batch of size $B$,

$$\mathbf{F}^{(m)} \in \mathbb{R}^{B \times d}, \quad \mathbf{F}^{(m)} = [\mathbf{f}_1^{(m)}, \dots, \mathbf{f}_B^{(m)}]^T.$$

The underlying multimodal alignment objective is the average of five terms,

$$L = \frac{1}{5}\left( L_{\text{CE}}^{\text{eeg-img}} + L_{\text{CE}}^{\text{img-eeg}} + L_{\text{CE}}^{\text{eeg-text}} + L_{\text{CE}}^{\text{text-eeg}} + L_{\cos} \right),$$

where the cross-entropy contrastive term is

$$L_{\text{CE}} = -\frac{1}{B} \sum_{i=1}^B \log \frac{ \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_i^{(m)}}{\tau}\right) }{ \sum_{j=1}^B \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_j^{(m)}}{\tau}\right) },$$

with $m \in \{\text{img}, \text{text}\}$, and the image–text consistency term is

$$L_{\cos} = 1 - \frac{1}{B} \sum_{i=1}^B \frac{\langle f_{\text{img}, i}, f_{\text{text}, i} \rangle} {\|f_{\text{img}, i}\| \|f_{\text{text}, i}\|}.$$

MCDB itself is not introduced as an additional loss. It is a gradient-level modulation scheme. The first step is to construct cross-modal similarity matrices between EEG and each other modality:

$$\mathbf{S}^{(m_1, m_2)} = \text{softmax}\left( \frac{\mathbf{F}^{(m_1)} (\mathbf{F}^{(m_2)})^T}{\tau} \right).$$

The method focuses on $m \in \{\text{eeg}, \text{img}, \text{text}\}$0 and $m \in \{\text{eeg}, \text{img}, \text{text}\}$1. Each matrix is summarized by summing its columns,

$m \in \{\text{eeg}, \text{img}, \text{text}\}$2

and the scalar contribution of modality $m \in \{\text{eeg}, \text{img}, \text{text}\}$3 is measured by $m \in \{\text{eeg}, \text{img}, \text{text}\}$4. The modality imbalance rate is then

$m \in \{\text{eeg}, \text{img}, \text{text}\}$5

This is converted into a soft suppression factor

$m \in \{\text{eeg}, \text{img}, \text{text}\}$6

where $m \in \{\text{eeg}, \text{img}, \text{text}\}$7 is a sensitivity coefficient. A dominant modality with $m \in \{\text{eeg}, \text{img}, \text{text}\}$8 receives $m \in \{\text{eeg}, \text{img}, \text{text}\}$9, while non-dominant modalities are left unchanged. Gradient descent is modified as

$$\mathbf{f}_i^{(m)} = \psi^{(m)}(\mathbf{x}_i^{(m)}; \boldsymbol{\theta}^{(m)}),$$0

The implementation is explicitly per-batch rather than per-sample. A single $$\mathbf{f}_i^{(m)} = \psi^{(m)}(\mathbf{x}_i^{(m)}; \boldsymbol{\theta}^{(m)}),$$1 and $$\mathbf{f}_i^{(m)} = \psi^{(m)}(\mathbf{x}_i^{(m)}; \boldsymbol{\theta}^{(m)}),$$2 are computed per modality for each iteration, and $$\mathbf{f}_i^{(m)} = \psi^{(m)}(\mathbf{x}_i^{(m)}; \boldsymbol{\theta}^{(m)}),$$3 is recomputed every iteration. The paper states that $$\mathbf{f}_i^{(m)} = \psi^{(m)}(\mathbf{x}_i^{(m)}; \boldsymbol{\theta}^{(m)}),$$4 for CLIP and reports the best performance at $$\mathbf{f}_i^{(m)} = \psi^{(m)}(\mathbf{x}_i^{(m)}; \boldsymbol{\theta}^{(m)}),$$5 in its ablation.

3. Architectural position and related realizations

The supplied papers place “modal consistency” and “dynamic balance” at different locations in the learning pipeline. Only HMAVD uses the exact term MCDB, but the other systems provide closely related realizations of the same design pattern.

Framework	Balance unit	Consistency/balance mechanism
HMAVD (Sun et al., 3 Sep 2025)	Per-batch optimization	EEG–image and EEG–text similarity drive $$\mathbf{f}_i^{(m)} = \psi^{(m)}(\mathbf{x}_i^{(m)}; \boldsymbol{\theta}^{(m)}),$$6 and $$\mathbf{f}_i^{(m)} = \psi^{(m)}(\mathbf{x}_i^{(m)}; \boldsymbol{\theta}^{(m)}),$$7
DMS (Tanaka et al., 15 Jun 2025)	Per-sample fusion	Confidence, uncertainty, and semantic consistency produce soft modality weights
SEIC (Li et al., 2 Aug 2025)	Per-cluster regularization	Image–text consistency plus EMA-weighted dynamic balance over cluster usage

In HMAVD, MCDB operates at the optimization level, not at feature fusion. It wraps the gradient updates of the modality-specific encoders after the batch loss and cross-modal similarity statistics have been computed. This makes it a training-time control mechanism.

Dynamic Modality Scheduling, by contrast, is a feature-level fusion framework for Multimodal Large Models. The supplied description explicitly characterizes DMS as an implementation of MCDB in the sense that modality weights are adjusted at the instance level according to confidence, uncertainty, and semantic consistency. Its core equations are

$$\mathbf{f}_i^{(m)} = \psi^{(m)}(\mathbf{x}_i^{(m)}; \boldsymbol{\theta}^{(m)}),$$8

$$\mathbf{f}_i^{(m)} = \psi^{(m)}(\mathbf{x}_i^{(m)}; \boldsymbol{\theta}^{(m)}),$$9

and

$B$0

DMS further introduces the Modality Weight Consistency Loss,

$B$1

with total objective $B$2. Here, dynamic balance is feature-weighting rather than gradient suppression, and modal consistency is represented by cosine-similarity-based semantic alignment (Tanaka et al., 15 Jun 2025).

Self-Enhanced Image Clustering offers a third realization. The paper does not use the exact term MCDB, but the supplied explanation characterizes its Stage 1 as an MCDB strategy built from cross-modal semantic consistency and a dynamic balancing regularizer. The consistency side is multi-level: instance-level contrast $B$3, cluster-assignment contrast $B$4, and cluster-center contrast $B$5. The balance side is

$B$6

where $B$7 is the batch-average assignment probability for cluster $B$8 in modality $B$9, and $$\mathbf{F}^{(m)} \in \mathbb{R}^{B \times d}, \quad \mathbf{F}^{(m)} = [\mathbf{f}_1^{(m)}, \dots, \mathbf{f}_B^{(m)}]^T.$$0 is an EMA-tracked historical assignment distribution. Here, dynamic balance acts over cluster usage rather than over modality gradients or fusion coefficients (Li et al., 2 Aug 2025).

The thesis on uncertainty-resilient multimodal learning via consistency-guided cross-modal transfer provides the broadest conceptual backdrop. Its abstract states that the central idea is to use cross-modal semantic consistency as a basis for robust representation learning by projecting heterogeneous modalities into a shared latent space, thereby mitigating modality gaps and supporting uncertainty estimation and stable feature learning. However, the supplied thesis material is only a skeleton; the chapters containing definitions of “dynamic balancing” and method details are absent, so no faithful MCDB mechanism can be reconstructed from it (Jang, 18 Nov 2025).

4. Empirical behavior across application domains

The HMAVD results provide the most direct empirical evidence for a named MCDB module. On the Extended ThingsEEG dataset, the task is zero-shot object recognition from EEG using 16,740 images, text labels, and EEG from 10 subjects. In the 63-channel setting, the full HMAVD model reports Top-1 accuracy of $$\mathbf{F}^{(m)} \in \mathbb{R}^{B \times d}, \quad \mathbf{F}^{(m)} = [\mathbf{f}_1^{(m)}, \dots, \mathbf{f}_B^{(m)}]^T.$$1 and Top-5 accuracy of $$\mathbf{F}^{(m)} \in \mathbb{R}^{B \times d}, \quad \mathbf{F}^{(m)} = [\mathbf{f}_1^{(m)}, \dots, \mathbf{f}_B^{(m)}]^T.$$2, compared with NICE at $$\mathbf{F}^{(m)} \in \mathbb{R}^{B \times d}, \quad \mathbf{F}^{(m)} = [\mathbf{f}_1^{(m)}, \dots, \mathbf{f}_B^{(m)}]^T.$$3 and $$\mathbf{F}^{(m)} \in \mathbb{R}^{B \times d}, \quad \mathbf{F}^{(m)} = [\mathbf{f}_1^{(m)}, \dots, \mathbf{f}_B^{(m)}]^T.$$4, respectively, yielding improvements of $$\mathbf{F}^{(m)} \in \mathbb{R}^{B \times d}, \quad \mathbf{F}^{(m)} = [\mathbf{f}_1^{(m)}, \dots, \mathbf{f}_B^{(m)}]^T.$$5 Top-1 and $$\mathbf{F}^{(m)} \in \mathbb{R}^{B \times d}, \quad \mathbf{F}^{(m)} = [\mathbf{f}_1^{(m)}, \dots, \mathbf{f}_B^{(m)}]^T.$$6 Top-5. The ablation on the sensitivity coefficient finds the best result at $$\mathbf{F}^{(m)} \in \mathbb{R}^{B \times d}, \quad \mathbf{F}^{(m)} = [\mathbf{f}_1^{(m)}, \dots, \mathbf{f}_B^{(m)}]^T.$$7, and the qualitative gradient analysis reports that after applying MCDB, text gradients are stabilized, image gradients are boosted toward the text level, and EEG gradients become more stable with less fluctuation. The additional runtime is also reported: NICE requires $$\mathbf{F}^{(m)} \in \mathbb{R}^{B \times d}, \quad \mathbf{F}^{(m)} = [\mathbf{f}_1^{(m)}, \dots, \mathbf{f}_B^{(m)}]^T.$$8 minutes, whereas HMAVD with MCDB and SPR requires $$\mathbf{F}^{(m)} \in \mathbb{R}^{B \times d}, \quad \mathbf{F}^{(m)} = [\mathbf{f}_1^{(m)}, \dots, \mathbf{f}_B^{(m)}]^T.$$9 minutes, or $$L = \frac{1}{5}\left( L_{\text{CE}}^{\text{eeg-img}} + L_{\text{CE}}^{\text{img-eeg}} + L_{\text{CE}}^{\text{eeg-text}} + L_{\text{CE}}^{\text{text-eeg}} + L_{\cos} \right),$$0 the time (Sun et al., 3 Sep 2025).

Dynamic Modality Scheduling supplies corresponding evidence in vision–language MLLMs. On clean data, DMS improves VQA accuracy from $$L = \frac{1}{5}\left( L_{\text{CE}}^{\text{eeg-img}} + L_{\text{CE}}^{\text{img-eeg}} + L_{\text{CE}}^{\text{eeg-text}} + L_{\text{CE}}^{\text{text-eeg}} + L_{\cos} \right),$$1 for static BLIP-2 to $$L = \frac{1}{5}\left( L_{\text{CE}}^{\text{eeg-img}} + L_{\text{CE}}^{\text{img-eeg}} + L_{\text{CE}}^{\text{eeg-text}} + L_{\text{CE}}^{\text{text-eeg}} + L_{\cos} \right),$$2, Recall@1 in retrieval from $$L = \frac{1}{5}\left( L_{\text{CE}}^{\text{eeg-img}} + L_{\text{CE}}^{\text{img-eeg}} + L_{\text{CE}}^{\text{eeg-text}} + L_{\text{CE}}^{\text{text-eeg}} + L_{\cos} \right),$$3 to $$L = \frac{1}{5}\left( L_{\text{CE}}^{\text{eeg-img}} + L_{\text{CE}}^{\text{img-eeg}} + L_{\text{CE}}^{\text{eeg-text}} + L_{\text{CE}}^{\text{text-eeg}} + L_{\cos} \right),$$4, and captioning CIDEr from $$L = \frac{1}{5}\left( L_{\text{CE}}^{\text{eeg-img}} + L_{\text{CE}}^{\text{img-eeg}} + L_{\text{CE}}^{\text{eeg-text}} + L_{\text{CE}}^{\text{text-eeg}} + L_{\cos} \right),$$5 to $$L = \frac{1}{5}\left( L_{\text{CE}}^{\text{eeg-img}} + L_{\text{CE}}^{\text{img-eeg}} + L_{\text{CE}}^{\text{eeg-text}} + L_{\text{CE}}^{\text{text-eeg}} + L_{\cos} \right),$$6. Under modality corruption, the gap widens: with image noise, static BLIP-2 falls to $$L = \frac{1}{5}\left( L_{\text{CE}}^{\text{eeg-img}} + L_{\text{CE}}^{\text{img-eeg}} + L_{\text{CE}}^{\text{eeg-text}} + L_{\text{CE}}^{\text{text-eeg}} + L_{\cos} \right),$$7 VQA accuracy while DMS reaches $$L = \frac{1}{5}\left( L_{\text{CE}}^{\text{eeg-img}} + L_{\text{CE}}^{\text{img-eeg}} + L_{\text{CE}}^{\text{eeg-text}} + L_{\text{CE}}^{\text{text-eeg}} + L_{\cos} \right),$$8; with text noise, static reaches $$L = \frac{1}{5}\left( L_{\text{CE}}^{\text{eeg-img}} + L_{\text{CE}}^{\text{img-eeg}} + L_{\text{CE}}^{\text{eeg-text}} + L_{\text{CE}}^{\text{text-eeg}} + L_{\cos} \right),$$9 while DMS reaches $$L_{\text{CE}} = -\frac{1}{B} \sum_{i=1}^B \log \frac{ \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_i^{(m)}}{\tau}\right) }{ \sum_{j=1}^B \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_j^{(m)}}{\tau}\right) },$$0. The ablation shows that removing confidence, uncertainty, or alignment degrades performance, with the largest retrieval drop occurring when alignment is removed, from $$L_{\text{CE}} = -\frac{1}{B} \sum_{i=1}^B \log \frac{ \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_i^{(m)}}{\tau}\right) }{ \sum_{j=1}^B \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_j^{(m)}}{\tau}\right) },$$1 Recall@1 to $$L_{\text{CE}} = -\frac{1}{B} \sum_{i=1}^B \log \frac{ \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_i^{(m)}}{\tau}\right) }{ \sum_{j=1}^B \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_j^{(m)}}{\tau}\right) },$$2. These results indicate that consistency-aware dynamic balancing is useful even when modalities are clean and increasingly important when they are corrupted, missing, or misaligned (Tanaka et al., 15 Jun 2025).

SEIC demonstrates the same principle in clustering rather than supervised multimodal prediction. On Tiny-ImageNet, removing any balance term yields ACC $$L_{\text{CE}} = -\frac{1}{B} \sum_{i=1}^B \log \frac{ \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_i^{(m)}}{\tau}\right) }{ \sum_{j=1}^B \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_j^{(m)}}{\tau}\right) },$$3, adding an “original balance term” yields $$L_{\text{CE}} = -\frac{1}{B} \sum_{i=1}^B \log \frac{ \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_i^{(m)}}{\tau}\right) }{ \sum_{j=1}^B \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_j^{(m)}}{\tau}\right) },$$4, and the proposed dynamic balance term reaches $$L_{\text{CE}} = -\frac{1}{B} \sum_{i=1}^B \log \frac{ \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_i^{(m)}}{\tau}\right) }{ \sum_{j=1}^B \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_j^{(m)}}{\tau}\right) },$$5. On CIFAR-100, the full SEIC configuration with a ViT-B/16 backbone reports ACC $$L_{\text{CE}} = -\frac{1}{B} \sum_{i=1}^B \log \frac{ \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_i^{(m)}}{\tau}\right) }{ \sum_{j=1}^B \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_j^{(m)}}{\tau}\right) },$$6, exceeding a cited prior result of $$L_{\text{CE}} = -\frac{1}{B} \sum_{i=1}^B \log \frac{ \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_i^{(m)}}{\tau}\right) }{ \sum_{j=1}^B \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_j^{(m)}}{\tau}\right) },$$7 obtained with a larger CLIP ViT-L/14 backbone. The ablation on consistency components further shows that the full combination of $$L_{\text{CE}} = -\frac{1}{B} \sum_{i=1}^B \log \frac{ \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_i^{(m)}}{\tau}\right) }{ \sum_{j=1}^B \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_j^{(m)}}{\tau}\right) },$$8, $$L_{\text{CE}} = -\frac{1}{B} \sum_{i=1}^B \log \frac{ \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_i^{(m)}}{\tau}\right) }{ \sum_{j=1}^B \exp\left(\frac{\mathbf{f}_i^{(\text{eeg})} \cdot \mathbf{f}_j^{(m)}}{\tau}\right) },$$9, and $m \in \{\text{img}, \text{text}\}$0 is stronger than reduced variants, especially on high-class settings such as Tiny-ImageNet (Li et al., 2 Aug 2025).

Across these settings, the empirical pattern is consistent. Dynamic balance is most useful when modality quality or representational strength is uneven, and modal consistency is most useful when semantic alignment must remain reliable despite noise, weak signals, or clustering degeneracy.

5. Interaction with adapters, uncertainty, and training stability

In HMAVD, MCDB is one component of a broader Dual Optimization Collaborative Mechanism. The other component is Stochastic Perturbation Regularization (SPR), while a separate Adapter module stabilizes high-dimensional image and text features before they enter the shared space. The adapter is described as a geometry-level or feature-level stabilization mechanism, whereas MCDB is an optimization-level balancing mechanism. The reported ablation shows that text alone slightly worsens Top-1 accuracy relative to the baseline, text plus adapter improves it, text plus MCDB plus SPR without the adapter does not improve beyond baseline, and the full model with text, adapter, MCDB, and SPR gives the best configuration. This indicates that balanced gradients, robust exploration, and feature stabilization are complementary rather than interchangeable (Sun et al., 3 Sep 2025).

DMS reveals a different interaction pattern. Its dynamic weights are computed from confidence, uncertainty, and semantic consistency, where confidence is estimated from predictive entropy, uncertainty from Monte Carlo dropout, and semantic consistency from cosine similarity to the average embedding of the other modalities. This makes the system instance-aware and quality-aware, but it also introduces the explicit cost of multiple stochastic forward passes for uncertainty estimation. The supplied discussion identifies several limitations: Monte Carlo dropout increases inference cost, the scheduler is heuristic because it uses a fixed linear combination with hand-chosen $m \in \{\text{img}, \text{text}\}$1, entropy is harder to summarize in generation tasks, pairwise one-vs-rest cosine similarity may be limited for richer multimodal structure, and the method assumes reasonably strong unimodal encoders (Tanaka et al., 15 Jun 2025).

SEIC highlights a further stability issue that arises when cross-modal consistency is combined with later encoder adaptation. Its Stage 2 no longer uses the Stage 1 cross-modal consistency losses because the text features are static while the visual encoder is then updated. Continuing to optimize the original alignment loss would force a drifting visual representation to align to fixed text features constructed for the old representation, and the supplied explanation reports collapse under that strategy. The method therefore switches to a confidence-weighted pseudo-label loss operating only on the visual branch during self-enhanced fine-tuning (Li et al., 2 Aug 2025).

For the broader thesis on consistency-guided cross-modal transfer, the available material states that the framework improves model stability, discriminative ability, and robustness to noisy or incomplete supervision on multimodal affect-recognition benchmarks, and that latent space analyses show reliable cross-modal structure under challenging conditions. At the same time, the missing method chapters mean that the actual architecture, equations, and any candidate dynamic-balancing rule cannot be truthfully specified from the supplied text (Jang, 18 Nov 2025).

6. Terminological ambiguity and unrelated acronym usage

The multimodal-learning use of MCDB is not yet a universally fixed label. The exact phrase “Modal Consistency Dynamic Balance” appears explicitly in the EEG–image–text decoding paper, while related work realizes the same conceptual pairing under different names such as Dynamic Modality Scheduling or dynamic balancing regularizer. Accordingly, the term should be read contextually: in one paper it is a per-batch gradient modulation scheme, in another it is a per-sample fusion scheduler, and in another it is a cluster-balance regularizer grounded in cross-modal consistency.

A separate and older literature uses the acronym MCDB in a wholly different sense. In chemical reaction network theory, MCDB denotes Markov chain detailed balance rather than Modal Consistency Dynamic Balance. There, the condition is the existence of a measure $m \in \{\text{img}, \text{text}\}$2 such that $m \in \{\text{img}, \text{text}\}$3 for all states of the continuous-time Markov chain. The paper proves that reaction network detailed balance is equivalent to Whittle stochastic detailed balance and implies Markov chain detailed balance, but the converse is false in general; it also provides an algorithm for finding conditions on rate constants required for MCDB in that sense (Joshi, 2013).

This terminological collision is a common source of confusion. In contemporary multimodal learning, MCDB refers to consistency-driven balancing among modalities. In stochastic chemical kinetics, MCDB refers to reversibility of a Markov chain. The two meanings are mathematically and historically unrelated, and disambiguation depends entirely on domain context.