Mixture-of-Geometry-Experts (MoGE) Overview

• Mixture-of-Geometry-Experts (MoGE) is a research direction integrating expert specialization, routing strategies, and geometry-preserving aggregation for advanced geometric inference.
• Implementations span dense monocular geometry estimators, sparse MoE methods for motion field decomposition, and novel aggregation operators tailored to embedding geometry.
• Empirical studies report strong performance across monocular, two-view, and 3D reconstruction tasks while highlighting trade-offs in architecture and expert decomposition.

Searching arXiv for recent papers directly relevant to “Mixture-of-Geometry-Experts (MoGE)” and closely adjacent formulations.
Mixture-of-Geometry-Experts (MoGE) does not yet denote a single standardized architecture in the recent literature. Instead, the phrase is most accurately treated as an umbrella over several neighboring developments: dense monocular geometry estimators whose name “MoGe” refers to monocular geometry estimation rather than mixture-of-experts; sparse Mixture-of-Experts systems applied directly to geometric prediction; geometry-preserving aggregation rules for expert outputs; mechanistic analyses showing that routers and experts co-organize geometrically; and geometry-aware expert decompositions of dynamical fields. Taken together, these works suggest that MoGE is less a settled model family than a research direction concerned with how expert specialization, routing, aggregation, and representation geometry should interact in geometric inference [2410.19115][2507.02546][2508.00592][2510.27234][2602.14039][2605.08648][2605.12476].

1. Terminological scope and competing uses of “MoGe”

The literature distinguishes sharply between MoGe as a proper model name and Mixture-of-Geometry-Experts as an interpretive label. In “MoGe: Unlocking Accurate Monocular Geometry Estimation for Open-Domain Images with Optimal Training Supervision” [2410.19115], “MoGe” names a monocular geometry estimator that predicts an affine-invariant 3D point map from a single image. The paper explicitly does not expand MoGe as “Mixture-of-Geometry-Experts.” “MoGe-2: Accurate Monocular Geometry with Metric Scale and Sharp Details” is a direct successor to that line and again centers affine-invariant monocular geometry rather than expert routing [2507.02546].

By contrast, several other papers are directly relevant to a MoGE interpretation even when they do not adopt the exact phrase. “GeoMoE: Divide-and-Conquer Motion Field Modeling with Mixture-of-Experts for Two-View Geometry” applies sparse MoE to heterogeneous motion sub-fields in two-view geometry [2508.00592]. “MoRE: 3D Visual Geometry Reconstruction Meets Mixture-of-Experts” inserts routed FFN experts into a dense 3D geometry transformer [2510.27234]. “Geometry-Preserving Aggregation for Mixture-of-Experts Embedding Models” studies whether the mixture rule itself should preserve the geometry of expert outputs [2602.14039]. “FLUX: Geometry-Aware Longitudinal Flow Matching with Mixture of Experts” combines a shared learned geometry with a mixture of expert vector fields [2605.08648]. “Routers Learn the Geometry of Their Experts: Geometric Coupling in Sparse Mixture-of-Experts” provides a mechanistic account of how standard sparse MoEs already partition hidden-state space geometrically [2605.12476].

This distribution of usage has an important consequence. The phrase “Mixture-of-Geometry-Experts” should not be read as naming one consensus blueprint. Rather, it designates a family resemblance across papers that differ in whether the “geometry” resides in the output representation, the routing decision, the aggregation operator, the learned manifold, or the expert specialization itself.

2. Affine-invariant monocular geometry: the MoGe and MoGe-2 lineage

MoGe formulates monocular geometry estimation (MGE) as dense 3D point-map prediction from a single RGB image (\mathbf I \in \mathbb R^{H\times W\times 3}), producing (\mathbf P \in \mathbb R^{H\times W\times 3}) in camera-related coordinates. Its central representational choice is an affine-invariant point map, meaning that the prediction is defined only up to a global scale and a global translation. This is motivated by focal-distance ambiguity in monocular reconstruction: scene size and focal length can trade off while leaving image evidence similar. The global geometry loss is
[
\mathcal L_G = \sum_{i\in \mathcal M}{1\over z_i}\left|s^*\hat{\mathbf p}_i+\mathbf t^*-\mathbf p_i\right|_1,
]
where ((s^*,\mathbf t^*)) are the optimal alignment parameters and (1/z_i) reweights the loss by inverse depth. MoGe further adds multi-scale local supervision on spherical regions
[
{\mathcal S}_j = { i \mid |{\mathbf p}_i - {\mathbf p}_j| \leq r_j,i\in {\mathcal M} },
]
and uses the ROE solver—described as robust, optimal, and efficient—to compute the alignment used in training. The method is trained on a mixed corpus of 21 public datasets and, in the reported zero-shot evaluation, achieves average affine-invariant point-map performance of Rel(^p = 6.43), (\delta_1^p = 94.4), average local point-map performance of Rel(^p = 5.33), (\delta_1^p = 95.8), and average camera FOV error of mean (2.91^\circ), median (2.21^\circ) [2410.19115].

MoGe-2 preserves this affine-invariant core and adds metric scale in a deliberately decoupled manner. The model uses a DINOv2 ViT-Large backbone, a shared convolutional neck, a dense geometry head that predicts the same affine-invariant point map as MoGe, and a separate global scale head. The preferred metric formulation is scale and relative geometry decomposition:
[
\mathcal{L}_{s} = | \text{log}(\hat{s}) - \mathop{\rm stopgrad}(\text{log}(s^*)) |_2^2,
]
where (\hat s) is predicted by a CLS-token-conditioned MLP, and (s^*) is the optimal online scale from the ROE solver. The stop-gradient enforces the decoupling: the scale target supervises the scale head without backpropagating through alignment into the geometry branch. The final metric point map is obtained by multiplying the affine-invariant point map by the predicted global scale. The paper explicitly states that this design is not a new expert mixture in the classical mixture-of-experts sense; architecturally it is a shared encoder plus dense geometry head plus global scale head [2507.02546].

MoGe-2 also introduces a unified real-data refinement pipeline motivated by the observation that LiDAR and SfM supervision contain noise, incompleteness, and RGB-depth misalignment that suppress sharp detail learning. A synthetic-only model (G_\text{syn}) provides sharp pseudo-geometry used for local mismatch filtering and logarithmic-space Poisson completion. The mismatch-filtering stage constructs local neighborhoods
[
\hat{\mathcal S}_j=\left{i \mid |\hat{\mathbf p}_i-\hat{\mathbf p}_j|\leq \hat r_j,i\in \mathcal M\right},
]
aligns real points locally to pseudo-predictions with the ROE solver, removes outliers, and then completes missing geometry by matching gradients of (\log \hat d_i) while preserving boundary depth. The resulting model is trained on 24 datasets and is evaluated along three axes: relative geometry, metric geometry, and boundary sharpness. Averaged over 10 datasets, it improves relative-geometry average rank from MoGe’s (2.53) to (2.05); averaged over 7 metric datasets, it reports metric point-map Rel(^p = 8.19), (\delta_1^p = 93.6), and average rank (1.95); for metric depth without GT intrinsics it reports Rel(^d = 15.7), (\delta_1^d = 76.8); and for boundary sharpness it ranks second overall behind Depth Pro, with average rank (1.75) [2507.02546].

Within a MoGE encyclopedia, the MoGe lineage is therefore foundational but terminologically atypical. Its importance lies in showing that geometry-aware representation design—especially affine invariance and alignment-aware supervision—can dominate architectural novelty. A plausible implication is that later expert-based systems inherit many of the same concerns: ambiguity management, scale decoupling, and local-versus-global geometric supervision.

3. Explicit Mixture-of-Experts formulations for geometric prediction

GeoMoE is among the clearest instances where a standard sparse MoE is used as a geometry model. It operates on two-view correspondences (\mathbf C={\mathbf c_i}\in\mathbb R^{N\times 4}), converts them into motion vectors (\mathbf m_i=(\mathbf x_i,\mathbf x'i-\mathbf x_i)), embeds them, and refines the resulting motion field through a layered pipeline comprising Local Orthogonal Context (LOC), Probabilistic Prior-Guided Decomposition (PPGD), MoE-Enhanced Bi-Path Rectifier (MBPR), graph-based reconstruction, and layer-wise inlier prediction. The MoE component uses sparse routing
[
\mathcal R=\mathrm{top\text{-}k}\big(\mathrm{softmax}(\mathrm{MLP}(\mathbf f_i))\big), \qquad
\mathbf f'_i=\sum{j\in \mathbb N_k}\mathcal R_j\cdot E_j(\mathbf f_i),
]
with 4 experts and 2 active experts per sub-field. The decomposition is prior-guided by previous-layer inlier probabilities, which act as soft masks during refinement. GeoMoE reports (L=8) layers, feature dimension (D=128), maximum (N=2000) correspondences, and (M=48) motion sub-fields. On YFCC100M it reports AUC values (34.21 / 56.25 / 73.53) at (5^\circ/10^\circ/20^\circ) using weighted eight-point estimation, and on SUN3D it reports (9.98 / 24.05 / 42.71). It also reports HPatches homography accuracy (53.83 / 64.92 / 74.60) with DLT at (3/5/10) px, 3DMatch registration recall (78.33), and an efficiency profile of (5.552)M parameters, (3.106)G FLOPs, (106.73)MB memory, and (56.14)ms inference time [2508.00592].

Conceptually, GeoMoE instantiates a divide-and-conquer view of geometry: the input correspondence field is treated as a mixture of latent motion regimes rather than as one globally smooth field. The experts are not explicitly named as “foreground,” “background,” or “planar,” but the paper states that visualizations show different experts responding to different objects, spatial regions, and characteristic motion patterns. This is a strong, though still emergent rather than hand-specified, MoGE interpretation.

MoRE extends sparse MoE into dense 3D visual geometry reconstruction from unposed image sequences ((I_i)^N_{i=1}), predicting ((C_i, P_i, D_i, T_i, N_i){i=1}^N) for camera parameters, pointmaps, depth, dense tracking features, and normals. It is built on VGGT, a dense visual transformer with alternating global and frame attention, and inserts MoE by replicating pretrained FFNs into expert sets ({\varepsilon_i}). Routing is standard sparse conditional computation:
[
\mathcal{P}(x)_i = \frac{e^{f(x)_i}}{\sum^E_j e^{f(x)_j}}, \qquad
\mathrm{MoE}(x) = \sum{i=1}^{K} \mathcal{P}(x)i \cdot \varepsilon(x)_i,
]
with a load-balancing objective
[
\mathcal{L}{\text{moe}} = E \cdot \sum_{i=1}^{E} \mathcal{F}_i \cdot \mathcal{G}_i.
]
The paper does not fully specify the number of experts (E), the number of MoE layers, or parameter/FLOP counts, but it is explicit that MoE is introduced in a second stage after initial dense training [2510.27234].

MoRE couples its MoE backbone to geometry-specific additions. A confidence-based depth refinement module uses MoGev2 as a teacher prior and defines a binary confidence mask
[
M_{\text{conf}} = \left{ \frac{|D_{\text{moge}} - D_{\text{gt}}|}{\max(D_{\text{gt}}, \alpha)} < \tau \right},
]
with (\alpha=0.5) and (\tau=0.1), to filter unreliable depth supervision. A normal-prediction branch fuses globally aligned 3D backbone features (f_{3d}) with dense semantic features (f_s) from DINOv2 via (f_n = f_{3d} \oplus f_s). Quantitatively, the paper attributes gains to both the tailored geometry losses and the MoE specialization: a VGGT-like baseline without MoE reports DTU Acc./Comp./N.C. (1.338/1.896/0.676), NYUv2 depth (0.056/0.951), and RealEstate10K AUC@30 (77.62); adding tailored losses but no MoE improves to (1.297/1.625/0.682), (0.054/0.953), and (85.14); full MoRE further improves to (1.011/1.491/0.695), (0.051/0.957), and (86.13) [2510.27234].

Across GeoMoE and MoRE, the salient MoGE pattern is explicit sparse routing inside a geometry pipeline. GeoMoE decomposes motion sub-fields before expert refinement; MoRE routes dense visual tokens inside a multi-task 3D backbone. In both cases, expert specialization is latent and data-driven rather than manually typed by geometry primitive.

4. Geometry-preserving aggregation and hyperspherical expert representations

A different branch of the literature asks whether MoGE should be defined not only by which experts are selected but by how their outputs are mixed. “Geometry-Preserving Aggregation for Mixture-of-Experts Embedding Models” studies sparse MoE embedding systems in which the conventional output is the weighted linear sum
[
y_{\text{linear}} = \sum_{i \in \text{Top-}K} w_i e_i .
]
The paper argues that this rule is geometrically mismatched for embedding models because the expert outputs empirically lie on a shared hyperspherical manifold: norms are tightly concentrated, while active experts are directionally separated, with most pairwise angles above (40^\circ). Under this geometry, linear mixing causes aggregation-induced inward collapse, formally expressed for approximately equal norms as
[
|y_{\text{linear}}| = r \left| \sum_i w_i \hat{e}_i \right| < r .
]
This degrades both norm and direction, and therefore embedding comparability [2602.14039].

The proposed remedy is Spherical Barycentric Aggregation (SBA), which decomposes each expert output into radius and direction,
[
r_i = |e_i|, \qquad \hat{u}i = \frac{e_i}{|e_i|},
]
aggregates radius linearly,
[
r = \sum{i=1}^{K} w_i r_i,
]
and aggregates angular coordinates with norm-aware weights,
[
\theta = \frac{\sum_{i=1}^{K} w_i r_i \theta_i}{\sum_{i=1}^{K} w_i r_i},
]
before reconstructing
[
y_{\text{SBA}} = r \cdot \hat{u}(\theta).
]
The paper is explicit that this is an operational geometry-preserving rule, not a rigorous intrinsic Fréchet mean on (S^{d-1}). It does not modify the experts or the router; it modifies only the aggregation operator [2602.14039].

Empirically, the paper compares a linear MoE baseline and SBA MoE on selected MTEB tasks using nomic-ai/nomic-embed-text-v2-moe with top-2 routing. Reported scores are (71.18 \rightarrow 71.51) on STSBenchmark, (85.66 \rightarrow 86.19) on StackExchangeClustering, and (83.66 \rightarrow 89.06) on SprintDuplicateQuestions. Ablations show that a norm-free angular variant performs similarly and that a unit-normalized output degrades strongly, including a reported (0.00) on StackExchangeClustering [2602.14039].

For a MoGE synthesis, the importance of this paper is architectural minimalism. It treats geometry-awareness as a property of the mixture operator rather than the expert modules themselves. This suggests that some MoGE gains may come from preserving the geometry of expert representations during aggregation, even when routing and expert internals are left unchanged.

5. Router–expert geometric coupling in sparse MoE

A mechanistic account of MoGE is developed in “Routers Learn the Geometry of Their Experts: Geometric Coupling in Sparse Mixture-of-Experts” [2605.12476]. The setting is a standard sparse MoE layer in which the router computes
[
\mathbf{p} = \sigma(\mathbf{z} + \mathbf{m}), \qquad \mathbf{z} = W_r\mathbf{x},
]
and the sparse output is
[
y = \sum_{i\in\mathcal{T}K} p_iE_i(\mathbf{x}).
]
The central result is that, for a routed token (\mathbf{x}), the selected router vector and the selected expert’s input-side weights receive gradients along the same input direction:
[
\nabla{\mathbf{w}{i,k}} \mathcal{L} = \delta{i,k}\,\mathbf{x}^{\top} \propto \mathbf{x}^{\top}, \qquad
\nabla_{\mathbf{r}_i} \mathcal{L} = \gamma_i \mathbf{x} \propto \mathbf{x}.
]
The scalar coefficients differ, but the directional component is shared. The paper’s interpretation is that matched router–expert directions accumulate the same routed token history [2605.12476].

This geometric-coupling view is supported empirically in a (1)B SMoE trained from scratch with (L=9) SMoE layers, hidden size (d=1024), (N=64) routed experts, top-(K=6) routing, (N_s=2) shared experts, and expert hidden width (512). For routed token–expert pairs, higher raw router scores correlate with stronger gate-neuron activations inside the selected expert, with pooled correlation
[
\rho = 0.43,\quad p = 1.2\times 10^{-81}.
]
The paper therefore argues that routing decisions are mirrored inside the selected expert rather than being merely external control signals [2605.12476].

The same paper also analyzes how auxiliary load-balancing losses alter this geometry. Under the standard balancing loss
[
\mathcal{L}{\mathrm{balance}} = N \sum{i=1}^{N} f_i\, P_i,
]
every router row receives input-directed updates from every token:
[
\nabla_{\mathbf{r}j}\mathcal{L}{\mathrm{balance}} = \beta_j\, \mathbf{x}, \qquad \beta_j \neq 0.
]
The reported effect is router homogenization: mean off-diagonal router-row cosine similarity rises to (0.63, 0.63, 0.57) at layers (0,4,8) with auxiliary loss, versus (0.32, 0.18, 0.13) under loss-free bias balancing. The paper characterizes this as making router directions “nearly three times more similar” [2605.12476].

To test whether routing can be reduced to explicit geometry, the authors introduce a parameter-free online K-Means router that scores experts by cosine similarity to running centroids:
[
s_i(\mathbf{x}) = \frac{\mathbf{c}_i^\top \mathbf{x}}{|\mathbf{c}_i|\,|\mathbf{x}|} + b_i,
]
with centroid update
[
\mathbf{c}_i \leftarrow \alpha\,\mathbf{c}_i + (1-\alpha)\,\bar{\mathbf{x}}_i.
]
This router achieves the lowest reported load imbalance, MaxVio (0.037), compared with (0.084) for Loss-Free, (0.102) for Loss-Free + Seq-Aux, and (0.526) for Aux-Loss, at the cost of modestly worse perplexity. In MoGE terms, this paper argues that sparse MoEs already behave like latent geometry-partitioning systems: experts occupy regions or directions in hidden-state space, and routers learn corresponding geometric descriptors.

6. Shared learned geometry with expert decomposition: FLUX

FLUX extends the MoGE discussion beyond vision and embeddings into longitudinal dynamical modeling. The problem is to infer a continuous-time velocity field from ordered but unpaired snapshot marginals (\mu_0,\ldots,\mu_{T-1}), with samples (\mathcal{D}k = {\mathbf{x}_k^{(i)}}{i=1}^{N_k}\subset\mathbb R^d). The model learns
[
v_\theta : [0,1]\times \mathbb{R}^d \to \mathbb{R}^d, \qquad \frac{d\mathbf{x}(t)}{dt} = v_\theta(t,\mathbf{x}(t)),
]
using adjacent-marginal flow matching. Geometry enters through a learned data-dependent metric, specifically an RBF-MLP scalar manifold score inducing an isotropic metric tensor
[
\mathbf{G}{\mathrm{RBF\text{-}MLP}}(\mathbf{x}) = M{\mathrm{RBF\text{-}MLP}}(\mathbf{x})\,\mathbf{I},
]
which is then used to construct geometry-aware conditional paths between adjacent marginals [2605.08648].

The expert decomposition acts on the velocity field rather than on the geometry itself:
[
v_\theta(t,\mathbf{x}) = \sum_{m=1}^{M} w_m(t,\mathbf{x}) f_m(t,\mathbf{x}),
]
where (f_m) are expert vector fields and (w_m) are produced by a Straight-Through Gumbel-Softmax router
[
w_m = \frac{ \exp((\ell_m+\eta_m)/\tau_g) }{ \sum_{q=1}^{M}\exp((\ell_q+\eta_q)/\tau_g) }, \qquad \eta_m \sim \mathrm{Gumbel}(0,1).
]
The training loss is the flow-matching regression
[
L_{\mathrm{FM}} = \mathbb{E} \left[ \left| v_\theta(t_{k,\alpha},\mathbf{z}{k,\alpha}) - \dot{\mathbf{z}}{k,\alpha} \right|_2^2 \right],
]
augmented by an extensive routing objective involving batch diversity, sparsity, temporal consistency, load balancing, confidence, DEC-style clustering, and segment-level regularization [2605.08648].

FLUX is especially informative because it isolates the role of shared geometry in enabling expert specialization. The paper explicitly states that the metric is not an additional input to the velocity network; instead, it changes the conditional paths, training locations, and target tangents used for supervision. Ablations compare full FLUX with “FLUX without Geometric Learning,” holding the MoE architecture and routing objective fixed while replacing geometry-aware paths with Euclidean ones. On the regime-switching Lorenz system, full FLUX reports ARI (=1.000), NMI (=1.000), whereas FLUX without geometry reports ARI (=0.000), NMI (=0.000). On NeuralTable, the corresponding values are (0.461/0.589) versus (0.200/0.235). On Embryoid Body, both variants report ARI (=0.571), NMI (=0.800), which the paper interprets as a case where stages are already spatially separable [2605.08648].

For MoGE, FLUX clarifies an important conceptual boundary. It is best described as a mixture of dynamics experts on a shared learned geometry, not as a system in which each expert owns a separate metric or manifold. The paper itself makes this distinction explicit: a stronger “true” MoGE would require per-expert geometries, per-expert path constructors, or routing over geometry families rather than only over vector fields.

7. Misconceptions, limits, and unresolved design questions

A recurrent misconception is to treat all occurrences of “MoGe” as references to mixture-of-experts geometry. That is incorrect. The MoGe and MoGe-2 papers are monocular geometry estimators built around affine-invariant point maps, optimal alignment, and decoupled metric-scale prediction; MoGe-2 explicitly notes that its scale head is not a separate “expert mixture” in the classical MoE sense [2410.19115][2507.02546]. Conversely, several papers that are central to a MoGE reading—GeoMoE, MoRE, SBA, FLUX, and router–expert coupling—do not define a unified canonical MoGE architecture [2508.00592][2510.27234][2602.14039][2605.08648][2605.12476].

A second misconception is that geometry-aware expert systems necessarily require geometry-specific expert modules. The literature is more heterogeneous. GeoMoE uses standard sparse experts but applies them to decomposed motion sub-fields [2508.00592]. MoRE uses FFN experts inside a dense 3D transformer without explicitly labeling them by geometry primitive [2510.27234]. SBA changes only the aggregation operator and leaves experts and router untouched [2602.14039]. The router-coupling work suggests that even a conventional sparse MoE already learns a geometric partition of hidden-state space [2605.12476]. FLUX uses a shared learned metric plus expert vector fields, but not per-expert metrics [2605.08648].

The current literature also leaves several technical gaps. GeoMoE does not provide the exact weighted eight-point formulation, the exact form of (\mathcal L_{reg}), the exact BCE expression, or the KNN neighborhood size (K) in the main text [2508.00592]. MoRE does not report the number of experts, top-(K), parameter counts, or FLOP overhead, and does not analyze what its experts learn [2510.27234]. SBA motivates hyperspherical consistency strongly but does not provide a rigorous spherical Fréchet mean or intrinsic manifold optimization formalism [2602.14039]. FLUX fixes the number of experts and trains geometry, bend network, and velocity network in separate stages, which reinforces shared geometry but limits expert-specific geometric adaptation [2605.08648]. MoGe-2 still struggles with extremely fine structures such as thin lines and hair, with preserving perfectly straight and aligned structures under large foreground–background scale differences, and with real-world metric-scale ambiguity in out-of-distribution settings [2507.02546].

These limitations indicate that MoGE remains a partially consolidated field. A plausible implication is that a more explicit next-generation formulation would combine several threads that are currently separate: affine- or metric-aware geometric representations; geometry-conditioned routing; geometry-preserving aggregation; mechanistically stable router–expert coupling; and, potentially, per-expert geometry rather than only per-expert function specialization. The present literature establishes the components of that program, but not yet a single definitive synthesis.