MoVa: A Multifaceted Research Acronym

• MoVa is a polysemous acronym used in various domains such as multimodal language modeling, video–audio generation, video–text alignment, and moral classification.
• Its implementations employ modular designs like mixture-of-vision experts, bidirectional diffusion transformers, and asymmetric dual projections to tackle different data modalities.
• Disambiguation is crucial since citations of MoVa may refer to distinct frameworks across machine learning, transportation engineering, and cryptography.

Searching arXiv for papers using the term “MoVa” / “MOVA” to ground the article in current research. MoVa, often stylized as MoVA or MOVA depending on the paper, is not a single unified concept but a recurring acronym used for several distinct research systems across machine learning, multimodal modeling, transportation engineering, cryptography, and computational social science. In recent arXiv literature, the name denotes at least four prominent contemporary frameworks: a multimodal LLM built around a mixture of vision experts, a joint video–audio generation model, a video–text alignment framework for long sequences, and a generalizable classifier of morals and values in text (Zong et al., 2024, Team et al., 9 Feb 2026, Zhu et al., 1 Jul 2026, Chen et al., 29 Sep 2025). Earlier and parallel uses include MOVA for adaptive traffic signal control (Cabrejas-Egea et al., 2021) and MOVA as the MOnnerat–VAudenay undeniable signature scheme (Meyer, 2011). The shared acronym therefore functions as a cross-domain label rather than a coherent technical lineage.

1. Multiplicity of the name

The term appears in arXiv as an overloaded acronym attached to unrelated technical objects. In multimodal learning, “MoVA: Adapting Mixture of Vision Experts to Multimodal Context” defines MoVA as a multimodal LLM that adaptively routes among heterogeneous vision encoders (Zong et al., 2024). In generative modeling, “MOVA: Towards Scalable and Synchronized Video-Audio Generation” uses MOVA for a large open-source joint audio–visual generator (Team et al., 9 Feb 2026). In video–language representation learning, “MoVA: Learning Asymmetric Dual Projections for Modular Long Video-Text Alignment” defines MoVA as a CLIP-style framework with dual asymmetric projections (Zhu et al., 1 Jul 2026). In computational social science, “MoVa: Towards Generalizable Classification of Human Morals and Values” uses MoVa for a benchmark-and-prompting suite for moral and value classification (Chen et al., 29 Sep 2025).

This dispersion is not merely orthographic. Each system uses a different expansion of the acronym and addresses a different problem class: multimodal perception, generative modeling, retrieval and alignment, or text classification. A plausible implication is that any technical discussion of “MoVa” requires immediate domain disambiguation.

Usage	Expansion in paper	Domain
MoVA	Mixture of Vision Experts Adapter	Multimodal LLMs
MOVA	MOSS Video and Audio	Joint video–audio generation
MoVA	Modular Long Video-Text Alignment	Video–text retrieval and representation learning
MoVa	Morals and Values	Moral and value classification in text
MOVA	Microprocessor Optimised Vehicle Actuation	Traffic signal control
MOVA	MOnnerat–VAudenay	Undeniable digital signatures

2. MoVA as a multimodal LLM

In “MoVA: Adapting Mixture of Vision Experts to Multimodal Context”, MoVA is a multimodal LLM that replaces the single-encoder visual front end typical of CLIP-derived MLLMs with a pool of task-specific vision experts (Zong et al., 2024). Its motivation is the empirical observation that no single vision encoder dominates across all image types: CLIP is strong on general image understanding but poor on document or chart content; Pix2Struct, Deplot, and Vary are stronger on text-heavy inputs; DINOv2, Co-DETR, and SAM are stronger on grounding, detection, or segmentation; and BiomedCLIP is stronger on medical imagery (Zong et al., 2024).

The architecture is organized as a coarse-to-fine mixture-of-vision-experts. In the coarse-grained stage, the LLM performs context-aware expert routing from the image, the user instruction, and natural-language descriptions of each expert’s capability. Routing is hard at the expert-selection level, with at most 3 experts activated per sample. In the fine-grained stage, the selected experts are fused by the Mixture-of-Vision-Expert Adapter (MoV-Adapter), a stack of $L=3$ adapter blocks comprising per-expert cross-attention, a dynamic multimodal gating network, and transformer refinement (Zong et al., 2024). The base encoder is CLIP ViT-L/336px; the paper also uses Vicuna-7B and Hermes-Yi-34B as LLM backbones, and a separate pretrained BERT text encoder supplies a [CLS] token for gating (Zong et al., 2024).

The routing supervision is generated offline by comparing the language modeling loss of $N+1$ proxy LLaVA-1.5-7B models, selecting expert $j$ for sample $i$ when $\mathcal{L}_{i,j} < \mathcal{L}_{i,0}$, then truncating to at most 3 experts (Zong et al., 2024). Training proceeds in three stages: MoV-Adapter pretraining on about 15M visual instruction samples, supervised finetuning on about 1.6M high-quality instruction examples, and expert-routing LoRA training using routing prompts and offline routing labels (Zong et al., 2024).

Empirically, the model is reported to outperform both single-expert and plain-fusion baselines across diverse benchmarks. For example, in the introductory comparison table, MoVA reaches MMBench 65.9, DocVQA 59.0, ChartQA 56.8, REC 86.4, RES 49.8, and SLAKE 66.3, exceeding both CLIP-only and naive concatenation baselines on those metrics (Zong et al., 2024). The larger MoVA-34B is reported at MMBench 81.3, MMBench-CN 79.0, MathVista 44.3, ChartQA 73.8, and DocVQA 84.2 (Zong et al., 2024). Ablations show that random routing, using all experts without routing, and removing MoV-Adapter each cause substantial degradation, especially on text-heavy tasks such as ChartQA and DocVQA (Zong et al., 2024). Within this line of work, MoVA denotes a vision-expert MoE over large pretrained encoders rather than an MoE internal to the LLM.

3. MOVA as a joint video–audio generation model

In “MOVA: Towards Scalable and Synchronized Video-Audio Generation”, MOVA stands for MOSS Video and Audio and denotes an open-source joint, synchronized video–audio generation model (Team et al., 9 Feb 2026). The model targets $p(V, A \mid c)$, where conditioning $c$ consists of text and optionally an initial image, and is motivated by the limitations of cascaded pipelines that generate video and audio separately. The paper argues that such pipelines suffer from error accumulation, unidirectional dependence, higher cost and latency, and weaker synchronization, especially for frame-level timing and lip sync (Team et al., 9 Feb 2026).

Architecturally, MOVA is a dual-tower Diffusion Transformer (DiT) with a bidirectional Bridge (Team et al., 9 Feb 2026). The video tower is Wan2.2 I2V A14B, the audio tower is a 1.3B DiT, and the Bridge contains 2.6B parameters. Total parameter count is 32B, with 18B active during inference because the video backbone uses Mixture-of-Experts FFN layers (Team et al., 9 Feb 2026). The model operates over video and audio VAEs and is trained with flow matching / rectified flow using a joint loss

$$\mathcal{L}_{FM} = \mathbb{E}_{t, c, x^v, x^a}\left[ \lambda_v \left\| v^v - \hat{v}^v_\theta(x^v_t, t, c)\right\|_2^2 + \lambda_a \left\| v^a - \hat{v}^a_\theta(x^a_t, t, c)\right\|_2^2 \right].$$

It further introduces Aligned RoPE to map video and audio tokens to a shared physical-time axis and dual sigma-shift schedules to stabilize multimodal denoising (Team et al., 9 Feb 2026).

MOVA supports IT2VA (Image-Text to Video-Audio) directly and exhibits T2VA (Text-only to Video-Audio) as an emergent capability by replacing the initial frame with a white image (Team et al., 9 Feb 2026). It is trained on >100,000 hours of video–audio data after extensive curation, including VAD, scene segmentation into 8.05s clips, audio quality filtering, DOVER video quality filtering, ImageBind or Synchformer alignment checks, and caption merging via large multimodal and LLMs (Team et al., 9 Feb 2026). Progressive joint training proceeds through a 360p baseline, a 360p quality-filtered alignment phase, and a 720p fine-tuning phase, using 1024 GPUs for a total of about 43,000 GPU-days (Team et al., 9 Feb 2026).

On Verse-Bench, the paper reports that MOVA-360p achieves IS = 4.269 and DNSMOS = 3.797, outperforming LTX-2 and OVI on those audio metrics (Team et al., 9 Feb 2026). With dual CFG, it achieves DeSync = 0.351, IB-Score = 0.315, LSE-D = 7.004, and LSE-C = 7.800, leading the listed joint open models on lip-sync metrics (Team et al., 9 Feb 2026). MOVA-720p attains cpCER = 0.149, better than LTX-2 and OVI on multi-speaker speech association (Team et al., 9 Feb 2026). Human evaluation yields ELO = 1113.8, above the compared open baselines (Team et al., 9 Feb 2026). In this usage, MOVA designates a large-scale generative AV backbone rather than a perception or alignment model.

A later watermarking paper, “mAVE: A Watermark for Joint Audio-Visual Generation Models”, uses MOVA-720p as a state-of-the-art evaluation target for cryptographic latent binding, characterizing MOVA as a native bimodal model with joint denoising, shared latent space, and Rectified Flow dynamics (Si et al., 7 Mar 2026). That paper is not a primary source for MOVA’s architecture, but it confirms the model’s role as a realistic open-source deployment target in joint audio–visual generation (Si et al., 7 Mar 2026).

4. MoVA as modular long video–text alignment

In “MoVA: Learning Asymmetric Dual Projections for Modular Long Video-Text Alignment”, MoVA denotes Modular Long Video-Text Alignment, a CLIP-style framework designed to address Temporal Misalignment and Semantic Asymmetry in long video–caption pairs (Zhu et al., 1 Jul 2026). The paper models a video as $\vV = (\vX_1,\ldots,\vX_T)$ and text as $\vT$, with shared latent representations $N+1$0 and $N+1$1 (Zhu et al., 1 Jul 2026). Its central claim is that monolithic global embeddings inherited from image–text CLIP are insufficient for long videos because captions typically align only to constrained temporal windows, while frame-level visual detail is richer and only partially specified by text (Zhu et al., 1 Jul 2026).

The formalism introduces dual sparse asymmetric projections through a text-side mask $N+1$2 and a video-side mask $N+1$3, constrained by

$N+1$4

The associated optimization minimizes the $N+1$5-size of the masks under a reconstruction-style constraint on masked encoder outputs (Zhu et al., 1 Jul 2026). The paper states an identification theorem showing that, under smooth invertibility, full joint support, temporal stability, and view diversity conditions, the underlying representation blocks are block-wise identifiable (Zhu et al., 1 Jul 2026). This theoretical development is unusual among CLIP-style retrieval papers and is central to the model’s positioning.

The practical architecture augments standard CLIP encoders with two modules: a Temporal Mask Network (TMN) on the text side and a Concept Mask Network (CMN) on the video side (Zhu et al., 1 Jul 2026). TMN is a two-layer Transformer that predicts frame-aware binary masks over text latent dimensions using a straight-through estimator; CMN is a single-layer Transformer plus attention pooling that predicts a visual mask conditioned on the selected text subset (Zhu et al., 1 Jul 2026). These modules operate alongside a standard global retrieval loss $N+1$6, yielding a total objective

$N+1$7

The model remains lightweight: 174.7M parameters, about 7.6% more than CLIP4Clip and fewer than ProST (Zhu et al., 1 Jul 2026).

MoVA is evaluated on MSVD, DiDeMo, ActivityNet, VideoUFO, and UltraVideo, and is also used as a text encoder in VideoCrafter2 for long-text-to-video generation (Zhu et al., 1 Jul 2026). On ActivityNet, it reports R@1 = 47.8 for text-to-video retrieval and R@1 = 46.7 for video-to-text retrieval, both rising further with DSL post-processing (Zhu et al., 1 Jul 2026). On DiDeMo, it reports R@1 = 57.5, and on long-caption datasets it reports R@1 = 62.4 on VideoUFO and 58.5 on UltraVideo, surpassing the listed baselines (Zhu et al., 1 Jul 2026). In the text-to-video generation setting, replacing CLIP or SmartCLIP with MoVA improves the VBench overall average to 0.8708, above the reported alternatives (Zhu et al., 1 Jul 2026). The paper’s qualitative analyses emphasize frame-specific word selection and concept-localized GradCAM, reinforcing the model’s claim to modular and interpretable alignment (Zhu et al., 1 Jul 2026).

5. MoVa as classification of morals and values

In “MoVa: Towards Generalizable Classification of Human Morals and Values”, MoVa stands for Morals and Values and refers to a benchmark-and-methods suite for generalizable classification of human morals and values in text (Chen et al., 29 Sep 2025). The contribution is explicitly tripartite: 16 labeled datasets and benchmarking results from four theoretically-grounded frameworks, a lightweight LLM prompting strategy, and a new application that helps evaluate psychological surveys (Chen et al., 29 Sep 2025). The four frameworks are Moral Foundations Theory, Schwartz’s Human Values, Morality-as-Cooperation, and Common Morality (Chen et al., 29 Sep 2025).

The methodological core is the all@once strategy, a prompt that asks an LLM to score all dimensions simultaneously in a structured JSON format rather than issuing separate binary prompts for each label (Chen et al., 29 Sep 2025). The paper recommends this strategy specifically, stating that it resembles the multi-label classifier chain formulation. It contrasts all@once with 1-by-1, MoVa + definition, MoVa + example, MoVa + reason, and several lexicon-enhanced variants (Chen et al., 29 Sep 2025). The paper also defines a probability extraction procedure from token logprobs and a lexicon–LLM fusion rule

$N+1$8

but its main practical recommendation remains plain all@once (Chen et al., 29 Sep 2025).

The benchmark spans in-domain and out-of-domain moral-foundations datasets such as MFTC, eMFD, and MFRC; new-domain MFT datasets including SC, MIC, ARG, VIG, and ValEval-MFT; new-framework datasets such as Webis-ArgValues-22, ValEval-Schwartz, ValueNet, MAC-D, and MoralChoice; and questionnaire-based applications to MFQ, PVQ, and MAQ (Chen et al., 29 Sep 2025). This breadth is central to the paper’s claim of generalizable classification.

Empirically, the paper reports that all@once substantially improves over 1-by-1 prompting on MFT, for example increasing Authority F1 on Reddit from 0.21 to 0.43 and Sanctity F1 from 0.14 to 0.38 with GPT‑4o‑mini (Chen et al., 29 Sep 2025). On out-of-domain MFT data, MoVa reportedly attains the highest AUC on all five foundations for SC, MIC, and VIG, and still remains close on ARG (Chen et al., 29 Sep 2025). It also performs competitively on the new MFT dimension Liberty/Oppression, with F1 on VIG reported at about 0.97 for MoVa (Chen et al., 29 Sep 2025). On Human Values, MoVa-DeepSeek is reported as strongest on several settings, while on ValEval-Schwartz both MoVa and MoVa-DeepSeek reach about 0.87 accuracy on the generalization set (Chen et al., 29 Sep 2025). In the survey-auditing application, MoVa flags multi-loaded items in MFQ, MAQ, and PVQ, suggesting a use beyond benchmark classification (Chen et al., 29 Sep 2025).

This usage of MoVa differs sharply from the multimodal or generative senses: it is neither a backbone nor an alignment module, but a framework-agnostic prompting and evaluation ecosystem for theory-laden text classification.

6. Earlier and parallel uses of MOVA

Before the current wave of MoVA-named machine learning systems, MOVA already had established meanings in transportation and cryptography. In traffic engineering, MOVA means Microprocessor Optimised Vehicle Actuation, a traffic-responsive, vehicle-actuated controller for single / isolated intersections (Cabrejas-Egea et al., 2021). The controller uses two induction loop detectors per lane, builds a virtual cell representation, estimates flow, degree of saturation, effective queue, and delay, and applies a stage decision rule based on a proprietary performance index $N+1$9 and threshold $j$0: terminate the stage if $j$1, otherwise extend it (Cabrejas-Egea et al., 2021). In the cited RL benchmarking paper, MOVA functions as a black-box commercial baseline and is contrasted with DDQN, Dueling DDQN, and A2C agents on several Vissim experiments (Cabrejas-Egea et al., 2021). The name therefore predates its recent reuse in multimodal ML.

In cryptography, MOVA is the MOnnerat–VAudenay undeniable signature scheme, an interactive digital signature system designed for extremely short signatures, such as 20 bits, at the cost of interactive verification with the signer or a server holding the secret key (Meyer, 2011). The SMS train-ticket paper instantiates MOVA with a Legendre-symbol-based homomorphism and uses it to build train tickets that fit within a 160-character SMS, with the signature encoded into about 4 printable characters (Meyer, 2011). This line of work is conceptually unrelated to the modern ML systems, but it demonstrates that the acronym had already been established as a technical term with a precise formal meaning.

A contemporary neighboring usage appears in “mAVE: A Watermark for Joint Audio-Visual Generation Models”, where MOVA-720p is not the watermarking method itself but one of the evaluated joint AV generation models (Si et al., 7 Mar 2026). That paper characterizes MOVA as a Rectified-Flow-based joint AV model with joint inversion and uses it to validate mAVE’s cross-modal binding defense (Si et al., 7 Mar 2026). This secondary usage reinforces the growing prominence of MOVA in generative audio–visual research.

7. Conceptual comparison and significance

Across these papers, the recurring acronym labels systems that share no common mechanism, but several patterns recur in how the name is used. First, recent ML instances of MoVA tend to denote modular or mixture-based adaptation layers rather than monolithic single-encoder designs. The multimodal MLLM MoVA uses multiple heterogeneous vision encoders with context-aware routing (Zong et al., 2024); the long video–text MoVA uses dual asymmetric projections to disentangle frame-conditioned semantics (Zhu et al., 1 Jul 2026); and the morals-and-values MoVa recommends all@once multilabel prompting that leverages inter-label dependence rather than isolated classification (Chen et al., 29 Sep 2025). This suggests a family resemblance at the level of design philosophy—adaptivity, modularity, and structured selection—even though the systems are formally unrelated.

Second, the most visible modern uses occur in multimodal research. One MoVA is a multimodal LLM integrating vision experts (Zong et al., 2024); another MOVA is a joint audio–visual generator with a large MoE video backbone and bidirectional bridge (Team et al., 9 Feb 2026); another MoVA is a video–text alignment framework for long sequences (Zhu et al., 1 Jul 2026). A plausible implication is that the name has become especially attractive in multimodal ML because it naturally accommodates expansions involving vision, video, values, or audio.

Third, the overload creates a genuine bibliographic ambiguity. A citation to “MoVA” without an arXiv identifier or full expansion can refer to markedly different technical objects: an MLLM (Zong et al., 2024), a retrieval model (Zhu et al., 1 Jul 2026), or a text-classification suite (Chen et al., 29 Sep 2025). A citation to “MOVA” can indicate a traffic controller (Cabrejas-Egea et al., 2021), an undeniable signature scheme (Meyer, 2011), or a joint AV generator (Team et al., 9 Feb 2026). For research communication, the practical consequence is that precise disambiguation by full title or identifier is necessary.

In encyclopedia terms, MoVa is best understood not as a single subject with one lineage, but as a polysemous research acronym whose most important current referents are a mixture-of-vision-experts MLLM (Zong et al., 2024), a scalable synchronized video–audio generator (Team et al., 9 Feb 2026), a modular long video–text alignment framework (Zhu et al., 1 Jul 2026), and a generalizable morals-and-values classification suite (Chen et al., 29 Sep 2025). Earlier and parallel uses in traffic control (Cabrejas-Egea et al., 2021) and cryptography (Meyer, 2011) remain historically important and continue to shape the acronym’s ambiguity.