VCDM: Multifaceted Research Frameworks

• VCDM is an overloaded acronym defining multiple research paradigms, including a type-II minimally modified gravity theory that replaces the cosmological constant with an auxiliary potential while retaining only the two tensorial degrees of freedom.
• In cosmology, VCDM supports dark energy reconstruction and early-universe models with applications ranging from smooth Hubble transitions to bounce cosmologies, often improving fits over ΛCDM.
• Beyond gravity, VCDM also denotes advanced models in machine learning—such as definition generation and visual commonsense discovery—and a particle physics construction for vector cold dark matter in hidden sectors.

VCDM is an overloaded acronym whose dominant meaning depends on disciplinary context. In gravitation and cosmology it most commonly denotes a type-II minimally modified gravity framework in which the cosmological constant is replaced by a potential $V(\phi)$ of an auxiliary, non-dynamical field while the theory continues to propagate only the two tensorial degrees of freedom of general relativity (Aoki et al., 2021). In other literatures, the same acronym denotes the “Variational Contextual Definition Modeler” for definition generation (Reid et al., 2020), a conceptual “Visual Commonsense Discovery Model” corresponding to the VCM architecture introduced for visual commonsense discovery (Shen et al., 2024), and a vector cold dark matter construction based on a $U(1)_X\times Z_2$ hidden sector (Farzan et al., 2012).

1. Nomenclature and scope

The acronym is not standardized across fields. In the gravity literature, closely related papers expand it as “$V(\phi)$ Cold Dark Matter,” “Vacuum Cold Dark Matter,” “Variable Cosmological Constant Dark Matter,” or “Vacuum Cosmology in Deformed Gravity” (Aoki et al., 2021). This suggests that, in practice, the acronym functions primarily as a label for a specific minimally modified gravity family rather than as a universally fixed expansion.

Usage	Expansion	Domain
VCDM	$V(\phi)$ Cold Dark Matter / related gravity labels	Modified gravity and cosmology
VCDM	Variational Contextual Definition Modeler	NLP, definition modeling
VCDM	Visual Commonsense Discovery Model	Vision-language modeling
VCDM	Vector Cold Dark Matter	Particle physics, dark matter

The gravitational usage is by far the most technically elaborated in the supplied literature, spanning formal Hamiltonian structure, cosmological reconstruction, compact objects, collapse, and early-universe phenomenology (Felice et al., 2022). The machine-learning and particle-physics usages are independent constructions that share only the acronym.

2. VCDM as a type-II minimally modified gravity theory

In its gravitational sense, VCDM is a type-II minimally modified gravity theory formulated in ADM variables with an auxiliary scalar $\phi$ and Lagrange multipliers enforcing constraints on the preferred foliation (Ganz et al., 2024). A representative unitary-gauge action is

$$S_{\mathrm{VCDM}}=\frac12\int d^4x\,N\sqrt{\gamma}\Big[ R + K_{ij}K^{ij}-K^2 - 2V(\varphi) -2\,\frac{\lambda^i}{N}\partial_i\varphi -\frac32\lambda_0^2 - 2\lambda_0\,(K+\varphi) \Big],$$

with $N$ the lapse, $\gamma_{ij}$ the spatial metric, $K_{ij}$ the extrinsic curvature, and $\lambda_0,\lambda^i$ auxiliary multipliers (Ganz et al., 2024). The defining property is that $U(1)_X\times Z_2$0 is auxiliary rather than dynamical, so the gravity sector retains only two propagating tensorial modes, with no propagating scalar graviton (Aoki et al., 2021).

A central structural result is that VCDM is not related to general relativity by a local field redefinition. The free transverse-traceless graviton sector is identical to GR, with

$U(1)_X\times Z_2$1

and hence $U(1)_X\times Z_2$2, but the nonlinear vacuum dynamics differ from GR, as shown by the mismatch of Bianchi-I shear dynamics under any type-Ia canonical transformation (Aoki et al., 2021). The phrase “non-uniqueness of the massless TT graviton” refers precisely to this fact: identical free graviton propagation does not uniquely determine the full vacuum theory to be GR (Aoki et al., 2021).

VCDM is closely related to cuscuton theory in unitary gauge. The two share the property of having only two local gravitational degrees of freedom and an elliptic, non-propagating “instantaneous” or “shadowy” mode. However, their solution spaces are not identical. All acceptable cuscuton solutions are VCDM solutions, but the converse fails: VCDM admits exact GR solutions, including Schwarzschild and Kerr in suitable foliations, that cuscuton cannot realize with a timelike scalar configuration (Felice et al., 2022). In this sense, VCDM is the broader class.

At the homogeneous level, the modified background can be written in effective-fluid form,

$U(1)_X\times Z_2$3

so the auxiliary sector behaves like a time-dependent dark-energy contribution without introducing a propagating scalar field (Mishra et al., 28 Mar 2026). The GR limit is recovered when $U(1)_X\times Z_2$4, in which case the modification reduces to a linear function of the mean curvature and contributes only a cosmological-constant term plus a boundary term (Aoki et al., 2021).

3. Late-time cosmology and dark-energy reconstruction

VCDM has been used extensively as a reconstruction framework for late-time acceleration. One influential phenomenological realization introduces a low-redshift tanh transition in the Hubble rate,

$U(1)_X\times Z_2$5

with data allowing transitions near $U(1)_X\times Z_2$6 or $U(1)_X\times Z_2$7 and improving the fit over $U(1)_X\times Z_2$8CDM by $U(1)_X\times Z_2$9 for the dataset considered (Felice et al., 2020). A later type-II MMG realization of the same VCDM-like behavior, constrained by CC, RSD, DESI BAO DR2, and Union3, found a smooth feature in $V(\phi)$0 centered at $V(\phi)$1 in the range $V(\phi)$2–$V(\phi)$3, with $V(\phi)$4 km s$V(\phi)$5 Mpc$V(\phi)$6, $V(\phi)$7, and a $V(\phi)$8 improvement over $V(\phi)$9CDM for the full dataset combination (Mishra et al., 28 Mar 2026).

A distinct branch of this literature embeds a sign-switching cosmological constant, $V(\phi)$0, into VCDM. In this construction, a mirror AdS–dS transition is realized through a smooth tanh or logistic profile in the auxiliary potential, yielding either an “agitated” transition with a bump in $V(\phi)$1 or a “quiescent” transition with a smooth slope change (Akarsu et al., 2024). When confronted with Planck, BAO, Pantheon+, and cosmic shear, $V(\phi)$2VCDM improved the fit over both $V(\phi)$3CDM and a GR-based $V(\phi)$4CDM implementation while alleviating both the $V(\phi)$5 and $V(\phi)$6 tensions; the favored transition redshift was $V(\phi)$7–$V(\phi)$8 depending on the dataset combination (Akarsu et al., 2024).

VCDM has also been used as a stable host for broad dark-energy parameterizations. A reconstruction program showed that any homogeneous and isotropic expansion history $V(\phi)$9 can be realized by an appropriate $\phi$0 without introducing fatal instabilities, including phantom behavior and phantom crossing (Arora et al., 5 Aug 2025). Within that framework, CPL, BA, JBP, EXP, and LOG parameterizations were tested against Planck 2018 and DESI BAO DR2, with the posteriors favoring $\phi$1 and $\phi$2, but with $\phi$3 km s$\phi$4 Mpc$\phi$5 and $\phi$6, so those particular realizations did not alleviate the major tensions (Arora et al., 5 Aug 2025). A Padé $\phi$7 cosmographic treatment instead found $\phi$8 and no robust evidence for the previously reported transition feature, indicating that current background data can force VCDM into a $\phi$9CDM-like limit when analyzed kinematically (Bhoi et al., 9 Jun 2026).

More recent data-driven VCDM studies have focused on smooth late-time transitions in $$S_{\mathrm{VCDM}}=\frac12\int d^4x\,N\sqrt{\gamma}\Big[ R + K_{ij}K^{ij}-K^2 - 2V(\varphi) -2\,\frac{\lambda^i}{N}\partial_i\varphi -\frac32\lambda_0^2 - 2\lambda_0\,(K+\varphi) \Big],$$0. A symbolic-regression program within VCDM identified the one-parameter form

$$S_{\mathrm{VCDM}}=\frac12\int d^4x\,N\sqrt{\gamma}\Big[ R + K_{ij}K^{ij}-K^2 - 2V(\varphi) -2\,\frac{\lambda^i}{N}\partial_i\varphi -\frac32\lambda_0^2 - 2\lambda_0\,(K+\varphi) \Big],$$1

which yields a transient accelerating and phantom phase without a future big-rip singularity and achieved stronger Bayesian evidence than $$S_{\mathrm{VCDM}}=\frac12\int d^4x\,N\sqrt{\gamma}\Big[ R + K_{ij}K^{ij}-K^2 - 2V(\varphi) -2\,\frac{\lambda^i}{N}\partial_i\varphi -\frac32\lambda_0^2 - 2\lambda_0\,(K+\varphi) \Big],$$2CDM for the datasets considered (Borghetto et al., 16 Jun 2026). A different phenomenological realization,

$$S_{\mathrm{VCDM}}=\frac12\int d^4x\,N\sqrt{\gamma}\Big[ R + K_{ij}K^{ij}-K^2 - 2V(\varphi) -2\,\frac{\lambda^i}{N}\partial_i\varphi -\frac32\lambda_0^2 - 2\lambda_0\,(K+\varphi) \Big],$$3

was studied jointly with neutrino extensions; in that setting, the preferred transition lay near $$S_{\mathrm{VCDM}}=\frac12\int d^4x\,N\sqrt{\gamma}\Big[ R + K_{ij}K^{ij}-K^2 - 2V(\varphi) -2\,\frac{\lambda^i}{N}\partial_i\varphi -\frac32\lambda_0^2 - 2\lambda_0\,(K+\varphi) \Big],$$4, and sterile-neutrino extensions could raise $$S_{\mathrm{VCDM}}=\frac12\int d^4x\,N\sqrt{\gamma}\Big[ R + K_{ij}K^{ij}-K^2 - 2V(\varphi) -2\,\frac{\lambda^i}{N}\partial_i\varphi -\frac32\lambda_0^2 - 2\lambda_0\,(K+\varphi) \Big],$$5 to $$S_{\mathrm{VCDM}}=\frac12\int d^4x\,N\sqrt{\gamma}\Big[ R + K_{ij}K^{ij}-K^2 - 2V(\varphi) -2\,\frac{\lambda^i}{N}\partial_i\varphi -\frac32\lambda_0^2 - 2\lambda_0\,(K+\varphi) \Big],$$6–$$S_{\mathrm{VCDM}}=\frac12\int d^4x\,N\sqrt{\gamma}\Big[ R + K_{ij}K^{ij}-K^2 - 2V(\varphi) -2\,\frac{\lambda^i}{N}\partial_i\varphi -\frac32\lambda_0^2 - 2\lambda_0\,(K+\varphi) \Big],$$7 km s$$S_{\mathrm{VCDM}}=\frac12\int d^4x\,N\sqrt{\gamma}\Big[ R + K_{ij}K^{ij}-K^2 - 2V(\varphi) -2\,\frac{\lambda^i}{N}\partial_i\varphi -\frac32\lambda_0^2 - 2\lambda_0\,(K+\varphi) \Big],$$8 Mpc$$S_{\mathrm{VCDM}}=\frac12\int d^4x\,N\sqrt{\gamma}\Big[ R + K_{ij}K^{ij}-K^2 - 2V(\varphi) -2\,\frac{\lambda^i}{N}\partial_i\varphi -\frac32\lambda_0^2 - 2\lambda_0\,(K+\varphi) \Big],$$9 while keeping $N$0 eV and $N$1 compatible with the Standard Model (Ladeira et al., 5 Jan 2026). An agnostic transition search then mapped into VCDM reported the strongest evidence in Planck+DESI DR2+DESY5, with a quintessence–phantom transition at

$N$2

exceeding $N$3, while explicitly concluding that such transitions do not resolve the $N$4 tension (Scherer et al., 29 Apr 2025). Across this literature, the impact on $N$5 is therefore model-dependent rather than universal.

4. Compact objects, gravitational collapse, and the early universe

In static, spherically symmetric stellar structure, VCDM can become observationally indistinguishable from GR. For barotropic perfect-fluid stars, imposing regularity at the center, $N$6, and physically motivated boundary conditions on the Misner–Sharp mass forces the VCDM and VCCDM equations to reduce exactly to the Tolman–Oppenheimer–Volkoff equations of GR, including the Schwarzschild–(de Sitter) exterior (Felice et al., 2021). Static stars therefore do not generically discriminate VCDM from GR.

Collapse problems reveal a sharper role for the preferred foliation. In spherical scalar collapse with asymptotically flat Minkowski slicing, the instantaneous mode can be integrated out and the equations reduce to GR in maximal slicing, but only in that physically preferred slicing (Jalali et al., 2023). For sufficiently large scalar-field amplitudes, an apparent horizon forms and the exterior relaxes to Schwarzschild in maximal slicing, whereas inside the horizon the lapse decreases toward zero, so the central singularity is never reached in finite preferred time (Jalali et al., 2023). An analogous dust-collapse analysis showed that Oppenheimer–Snyder collapse in VCDM corresponds to a constant-$N$7 foliation of the GR solution, again ending in a static configuration with vanishing lapse at a radius inside the apparent horizon and again implying the need for a UV completion inside the black hole while leaving the exterior evolution under control for all finite cosmological times (Felice et al., 2022).

Early-universe applications exploit the same absence of extra propagating gravitational degrees of freedom. A bouncing cosmology was constructed within VCDM using a k-essence matter sector and an auxiliary potential engineered to realize an asymmetric bounce (Ganz et al., 2022). The scenario is reported as free of ghost and gradient instabilities, free of ad hoc matching conditions, and free of the anisotropic-stress BKL problem, while yielding an almost scale-invariant scalar spectrum after appropriate choices of the equation of state and the potential (Ganz et al., 2022). The same framework was extended to cubic interactions and the bispectrum: if the VCDM potential does not affect slow-roll inflaton dynamics, the leading bispectrum reproduces the standard single-field result, including Maldacena’s squeezed-limit consistency relation; if the VCDM potential dominates, the squeezed bispectrum can instead violate that relation and generate order-unity local non-Gaussianity (Ganz et al., 2024). Applied to the VCDM bounce, the predicted local $N$8 remains within current Planck bounds, while the tensor sector prefers a blue spectrum to satisfy CMB limits on $N$9 (Ganz et al., 2024).

5. Machine-learning uses: definition modeling and visual commonsense

In natural-language processing, VCDM denotes the “Variational Contextual Definition Modeler,” a conditional variational autoencoder for contextual definition generation (Reid et al., 2020). The model uses a bi-encoding architecture: a BERT-base context encoder processes a phrase–context pair, a second BERT-base encoder reads the gold definition during training, and a continuous latent variable $\gamma_{ij}$0 mediates between phrase semantics and generated definition. The generative formulation is

$\gamma_{ij}$1

with an ELBO objective and KL-annealing plus free bits to avoid posterior collapse (Reid et al., 2020). The decoder is a single-layer LSTM conditioned on $\gamma_{ij}$2 and attentively reading contextual encoder states. On Oxford, Urban, Wikipedia, Cambridge, and the French Robert dataset, the model achieved state-of-the-art or near-state-of-the-art automatic and human evaluation scores, including BLEU $\gamma_{ij}$3 and BERTScore F1 $\gamma_{ij}$4 on Oxford and BLEU $\gamma_{ij}$5 on Robert (Reid et al., 2020).

In computer vision, the 2024 paper on visual commonsense discovery introduces the task of VCD and a model officially named VCM; the accompanying details describe VCM as conceptually exactly a “Visual Commonsense Discovery Model (VCDM),” although the published acronym in the paper itself is VCM (Shen et al., 2024). The associated Visual Commonsense Dataset contains 106,277 images, 2,449,126 bounding boxes, 18,136 unique object names, and over 14 million object–commonsense pairs (Shen et al., 2024). Commonsense is represented as grounded triples

$\gamma_{ij}$6

where the head object is linked to an image bounding box and the relation $\gamma_{ij}$7 belongs to a taxonomy that distinguishes explicit/seen from implicit/unseen knowledge across property, action, and space categories (Shen et al., 2024). The OFA-based model receives the image, object name, ROI coordinates, and the requested commonsense type, and autoregressively emits multiple tail phrases separated by $\gamma_{ij}$8. Its best configuration, VCM$\gamma_{ij}$9, reaches BLEU-1 $K_{ij}$0, BLEU-2 $K_{ij}$1, ROUGE-L $K_{ij}$2, METEOR $K_{ij}$3, and WER $K_{ij}$4, and improves downstream performance on ImageNetVC and VQAv2 relative to the OFA baselines (Shen et al., 2024).

These two machine-learning usages are unrelated except for the acronym. One is a latent-variable sequence model for lexical semantics, the other a grounded vision-language generator trained for object-specific commonsense discovery.

6. Particle-physics usage: vector cold dark matter

In particle-physics summaries, VCDM can denote vector cold dark matter in a hidden-sector $K_{ij}$5 construction (Farzan et al., 2012). The gauge group is

$K_{ij}$6

with Standard Model fields neutral under $K_{ij}$7, a new gauge boson $K_{ij}$8, and one or two complex scalar singlets charged under $K_{ij}$9 (Farzan et al., 2012). The discrete $\lambda_0,\lambda^i$0 symmetry renders the vector odd and forbids kinetic mixing with hypercharge, making the vector stable after spontaneous symmetry breaking (Farzan et al., 2012).

In the minimal realization with one complex scalar $\lambda_0,\lambda^i$1, the potential is

$\lambda_0,\lambda^i$2

and the vector mass is

$\lambda_0,\lambda^i$3

with $\lambda_0,\lambda^i$4 the singlet vacuum expectation value (Farzan et al., 2012). The theory then couples to the Standard Model exclusively through Higgs-portal mixing. The two-scalar extension is richer: depending on the vacuum, it can exhibit spontaneous CP violation or an accidental $\lambda_0,\lambda^i$5 symmetry that stabilizes one scalar in addition to the vector, producing multicomponent dark matter (Farzan et al., 2012).

Relic abundance proceeds either through Higgs-portal annihilation into Standard Model states or through annihilation into dark scalars when kinematically open. Direct detection is spin-independent and Higgs mediated; the paper’s representative parameter choices give cross sections around $\lambda_0,\lambda^i$6–$\lambda_0,\lambda^i$7 cm$\lambda_0,\lambda^i$8, below the XENON100 bounds quoted in that work (Farzan et al., 2012). Collider phenomenology is correspondingly dominated by modified Higgs rates, possible invisible Higgs decays when dark scalars are light, and additional Higgs-like singlet states with suppressed production cross sections (Farzan et al., 2012).

Taken together, these usages show that “VCDM” is not a single concept but a family of acronymic conventions spanning modified gravity, cosmological dark-energy modeling, NLP, vision-language systems, and hidden-sector dark matter. The gravitational usage is the most expansive and internally coherent: it defines a type-II minimally modified gravity framework with two tensorial degrees of freedom, a preferred foliation, and a broad reconstruction space for late-time acceleration, compact-object dynamics, and early-universe cosmology (Aoki et al., 2021).