CorrMoE: Disambiguation & Key Advances

• CorrMoE is an overloaded term that encompasses a normalized optimal-transport dependence measure, a computer vision correspondence pruning framework, and a correlation-aware Mixture-of-Experts for CTR prediction.
• Each variant employs specialized methodologies—optimal transport and Gini-based normalization, dual-branch fusion with FlowAttention, and explicit cross-expert de-correlation—to address redundancy and dependence.
• These distinct implementations highlight the need for contextual interpretation, careful computational trade-offs, and parameter tuning to effectively manage correlation in diverse applications.

CorrMoE is an overloaded term in recent arXiv literature. In one usage, it denotes the Earth Mover’s Correlation, abbreviated in the underlying paper as eCor, a dependence measure defined on arbitrary metric spaces through the Wasserstein-1 distance between a joint distribution and the product of its marginals (Móri et al., 2020). In a second usage, it denotes a computer-vision architecture, “CorrMoE: Mixture of Experts with De-stylization Learning for Cross-Scene and Cross-Domain Correspondence Pruning,” designed for robust two-view correspondence pruning under domain shift and scene diversity (Xia et al., 16 Jul 2025). In a third, broader usage, it denotes a correlation-aware Mixture-of-Experts principle instantiated as De-Correlated MoE (D-MoE) for click-through-rate prediction, in which cross-expert correlation is explicitly measured and minimized during training (Wang et al., 23 May 2025). These usages are conceptually unrelated at the object level—statistical dependence, visual correspondence pruning, and recommendation modeling—but they share a common concern with correlation structure.

1. Terminological scope and disambiguation

The label “CorrMoE” is used in three distinct senses in the supplied literature. The earliest is a dependence measure on metric spaces, introduced as the Earth Mover’s Correlation and denoted by eCor; the exposition explicitly identifies it as “CorrMoE” (Móri et al., 2020). A later use designates a vision framework for correspondence pruning, where the name expands to “Mixture of Experts with De-stylization Learning for Cross-Scene and Cross-Domain Correspondence Pruning” (Xia et al., 16 Jul 2025). A third use treats CorrMoE as a general design principle for Mixture-of-Experts models that are “explicitly correlation-aware,” with D-MoE presented as an implementation for CTR prediction (Wang et al., 23 May 2025).

This multiplicity of meanings is not merely lexical. In the Earth Mover’s Correlation, “correlation” refers to statistical dependence between random variables. In the CTR formulation, correlation refers to redundancy across expert outputs inside an MoE ensemble. In the vision formulation, the string “Corr” is tied to correspondence pruning rather than to a formal correlation functional. A plausible implication is that the term should be interpreted contextually rather than as a stable technical acronym.

2. CorrMoE as the Earth Mover’s Correlation

In “The Earth Mover’s Correlation,” CorrMoE is defined as a normalized optimal-transport dependence measure on arbitrary metric spaces (Móri et al., 2020). Let $(X\text{-space}, \delta_X)$ and $(Y\text{-space}, \delta_Y)$ be metric spaces, and let the product space use an admissible metric $d$, with a typical choice the Manhattan metric

$d[(x,y),(u,v)] = \delta_X(x,u) + \delta_Y(y,v).$

The earth mover’s covariance is

$\eCov(X,Y) := e\big(P_{XY},\, P_X \otimes P_Y\big),$

that is, the earth mover’s distance between the joint law and the product of marginals. The earth mover’s variance is

$\eVar(X) := \eCov(X,X) = e\big(P_{XX},\, P_X \otimes P_X\big),$

and the paper proves that

$\eVar(X) = \mathbb{E}\big[\delta_X(X, X')\big],$

so the normalization is based on Gini’s mean difference. The resulting earth mover’s correlation is

$\eCor(X,Y) := \frac{\eCov(X,Y)}{\min\big\{\eVar(X),\, \eVar(Y)\big\}},$

provided the denominator is nonzero.

The construction was motivated by gaps in existing dependence measures. Pearson’s correlation does not characterize independence in general, and distance correlation, although strong in Euclidean and certain metric settings, depends on negative-type structure or Hilbert embeddability. CorrMoE addresses arbitrary metric spaces, including settings such as arbitrary graphs or mixed structured spaces, where those assumptions may fail (Móri et al., 2020).

The paper’s axiomatic discussion is central. In general metric spaces, CorrMoE satisfies independence characterization and continuity, and it proves weaker forms of similarity invariance and functional attainment at $1$. Specifically, $\eCor(X,Y)=0$ iff $(Y\text{-space}, \delta_Y)$0 and $(Y\text{-space}, \delta_Y)$1 are independent; continuity follows under weak convergence with the stated moment conditions; and if $(Y\text{-space}, \delta_Y)$2 almost surely for a similarity $(Y\text{-space}, \delta_Y)$3, then $(Y\text{-space}, \delta_Y)$4. Full separate-similarity invariance and the converse part of the attainment axiom are conjectured in Banach spaces but not proved in general metric spaces (Móri et al., 2020).

The bounding inequalities establish the normalization:

$(Y\text{-space}, \delta_Y)$5

hence $(Y\text{-space}, \delta_Y)$6. Symmetry in $(Y\text{-space}, \delta_Y)$7 and $(Y\text{-space}, \delta_Y)$8 holds by construction. The paper also gives a multivariate extension, defining $(Y\text{-space}, \delta_Y)$9 via the optimal transport cost between the joint law and the product of three marginals, then normalizing by $d$0.

Several special cases are explicit. For binary indicators,

$d$1

$d$2

and

$d$3

For bivariate normal $d$4 with correlation $d$5,

$d$6

with strict inequality unless $d$7, and the authors conjecture equality in the first inequality. For real-valued $d$8 with cdf $d$9,

$d[(x,y),(u,v)] = \delta_X(x,u) + \delta_Y(y,v).$0

3. Estimation and computation for the Earth Mover’s Correlation

The empirical version of CorrMoE is formulated as an optimal transport problem between two discrete measures on $d[(x,y),(u,v)] = \delta_X(x,u) + \delta_Y(y,v).$1 (Móri et al., 2020). Given observations $d[(x,y),(u,v)] = \delta_X(x,u) + \delta_Y(y,v).$2, one empirical measure places mass $d[(x,y),(u,v)] = \delta_X(x,u) + \delta_Y(y,v).$3 on each observed pair, while the other places mass $d[(x,y),(u,v)] = \delta_X(x,u) + \delta_Y(y,v).$4 on each grid point $d[(x,y),(u,v)] = \delta_X(x,u) + \delta_Y(y,v).$5. The empirical earth mover’s covariance is the minimal transport cost between these measures under the Manhattan product metric, with cost matrix

$d[(x,y),(u,v)] = \delta_X(x,u) + \delta_Y(y,v).$6

The empirical variance is the average pairwise distance,

$d[(x,y),(u,v)] = \delta_X(x,u) + \delta_Y(y,v).$7

and the empirical correlation is the ratio

$d[(x,y),(u,v)] = \delta_X(x,u) + \delta_Y(y,v).$8

Consistency is stated under standard assumptions ensuring convergence in $d[(x,y),(u,v)] = \delta_X(x,u) + \delta_Y(y,v).$9: empirical measures converge to their population counterparts, implying $\eCov(X,Y) := e\big(P_{XY},\, P_X \otimes P_Y\big),$0 and $\eCov(X,Y) := e\big(P_{XY},\, P_X \otimes P_Y\big),$1 when the limit denominator is nonzero. No closed-form asymptotic distribution for $\eCov(X,Y) := e\big(P_{XY},\, P_X \otimes P_Y\big),$2 is provided; for independence testing, the paper indicates that using $\eCov(X,Y) := e\big(P_{XY},\, P_X \otimes P_Y\big),$3 with permutation or bootstrap calibration is practical (Móri et al., 2020).

Computation is substantially more demanding than for distance correlation. The exact linear-programming formulation is a transportation problem with $\eCov(X,Y) := e\big(P_{XY},\, P_X \otimes P_Y\big),$4 supply nodes and $\eCov(X,Y) := e\big(P_{XY},\, P_X \otimes P_Y\big),$5 demand nodes. The exposition notes that naive assignment-style formulations can scale up to $\eCov(X,Y) := e\big(P_{XY},\, P_X \otimes P_Y\big),$6, while empirical eCov is generally more computationally intensive than dCor, which is $\eCov(X,Y) := e\big(P_{XY},\, P_X \otimes P_Y\big),$7. For scalable approximation, entropic regularization via Sinkhorn is proposed: form the kernel $\eCov(X,Y) := e\big(P_{XY},\, P_X \otimes P_Y\big),$8, solve for scaling vectors matching the marginals, and recover an approximate transport plan. The text also notes implementation details such as on-the-fly construction of the cost matrix and exploiting the separability of $\eCov(X,Y) := e\big(P_{XY},\, P_X \otimes P_Y\big),$9 to avoid storing $\eVar(X) := \eCov(X,X) = e\big(P_{XX},\, P_X \otimes P_X\big),$0 entries (Móri et al., 2020).

A notable caveat arises in finite samples: even if $\eVar(X) := \eCov(X,X) = e\big(P_{XX},\, P_X \otimes P_X\big),$1 is a one-to-one function of $\eVar(X) := \eCov(X,X) = e\big(P_{XX},\, P_X \otimes P_X\big),$2 in a discrete sample, empirical $\eVar(X) := \eCov(X,X) = e\big(P_{XX},\, P_X \otimes P_X\big),$3 need not equal $\eVar(X) := \eCov(X,X) = e\big(P_{XX},\, P_X \otimes P_X\big),$4. The example $\eVar(X) := \eCov(X,X) = e\big(P_{XX},\, P_X \otimes P_X\big),$5 has $\eVar(X) := \eCov(X,X) = e\big(P_{XX},\, P_X \otimes P_X\big),$6 and $\eVar(X) := \eCov(X,X) = e\big(P_{XX},\, P_X \otimes P_X\big),$7 (Móri et al., 2020). This directly qualifies a common misconception that perfect sample bijectivity must force maximal empirical dependence under every normalized statistic.

4. CorrMoE as correspondence pruning in computer vision

In computer vision, CorrMoE denotes a correspondence pruning framework for two-view geometry under cross-domain and cross-scene variation (Xia et al., 16 Jul 2025). The task is to assign each putative correspondence a probability of being geometrically consistent and to reject outliers before estimating the essential matrix. The input is a set of correspondences

$\eVar(X) := \eCov(X,X) = e\big(P_{XX},\, P_X \otimes P_X\big),$8

with

$\eVar(X) := \eCov(X,X) = e\big(P_{XX},\, P_X \otimes P_X\big),$9

normalized by camera intrinsics. A pruning model outputs scores $\eVar(X) = \mathbb{E}\big[\delta_X(X, X')\big],$0 or weights $\eVar(X) = \mathbb{E}\big[\delta_X(X, X')\big],$1 used in downstream weighted geometry solvers.

The architecture is organized around two named components. The De-stylization Dual Branch (DDB) addresses cross-domain shift by applying Progressive MixStyle-based style mixing to both implicit and explicit graph features. The Bi-Fusion Mixture of Experts (BF-MoE) addresses scene diversity through multi-expert fusion with linear-complexity attention and dynamic top-$\eVar(X) = \mathbb{E}\big[\delta_X(X, X')\big],$2 routing. The system-level flow is: input correspondences and initial node features via an MLP embedding; DDB with a parallel implicit branch and explicit branch; BF-MoE with FlowAttention, gating, expert routing, and residual normalization; a prediction head producing pruning weights; then weighted essential-matrix estimation and full-size verification (Xia et al., 16 Jul 2025).

The input graph representation separates implicit and explicit locality. In the implicit branch, neighborhoods are captured through DiffPool, OA filter, and DiffUnpool without explicit adjacency. In the explicit branch, a KNN graph is built in feature space, with edge features

$\eVar(X) = \mathbb{E}\big[\delta_X(X, X')\big],$3

followed by multi-dimensional attention across channel, neighbor, and spatial dimensions, and aggregation by AnnualConv. The DDB applies PMix before branch processing and again after branch reconstruction. The PMix schedule uses a linearly increasing application probability,

$\eVar(X) = \mathbb{E}\big[\delta_X(X, X')\big],$4

For a batch $\eVar(X) = \mathbb{E}\big[\delta_X(X, X')\big],$5, after computing per-channel statistics and shuffling to $\eVar(X) = \mathbb{E}\big[\delta_X(X, X')\big],$6, the mixed statistics are

$\eVar(X) = \mathbb{E}\big[\delta_X(X, X')\big],$7

with $\eVar(X) = \mathbb{E}\big[\delta_X(X, X')\big],$8, and the de-stylized feature is

$\eVar(X) = \mathbb{E}\big[\delta_X(X, X')\big],$9

The BF-MoE module first concatenates implicit and explicit features, then applies FlowAttention rather than quadratic softmax attention. The paper states that standard attention has complexity $\eCor(X,Y) := \frac{\eCov(X,Y)}{\min\big\{\eVar(X),\, \eVar(Y)\big\}},$0, while FlowAttention yields near-linear scaling with respect to sequence length, described at a high level by

$\eCor(X,Y) := \frac{\eCov(X,Y)}{\min\big\{\eVar(X),\, \eVar(Y)\big\}},$1

with complexity becoming $\eCor(X,Y) := \frac{\eCov(X,Y)}{\min\big\{\eVar(X),\, \eVar(Y)\big\}},$2 rather than $\eCor(X,Y) := \frac{\eCov(X,Y)}{\min\big\{\eVar(X),\, \eVar(Y)\big\}},$3. Routing is produced by

$\eCor(X,Y) := \frac{\eCov(X,Y)}{\min\big\{\eVar(X),\, \eVar(Y)\big\}},$4

and expert fusion is

$\eCor(X,Y) := \frac{\eCov(X,Y)}{\min\big\{\eVar(X),\, \eVar(Y)\big\}},$5

The reported default stacks $\eCor(X,Y) := \frac{\eCov(X,Y)}{\min\big\{\eVar(X),\, \eVar(Y)\big\}},$6 MoE layers, each with $\eCor(X,Y) := \frac{\eCov(X,Y)}{\min\big\{\eVar(X),\, \eVar(Y)\big\}},$7 experts and top-$\eCor(X,Y) := \frac{\eCov(X,Y)}{\min\big\{\eVar(X),\, \eVar(Y)\big\}},$8 routing (Xia et al., 16 Jul 2025).

Training uses the combined objective

$\eCor(X,Y) := \frac{\eCov(X,Y)}{\min\big\{\eVar(X),\, \eVar(Y)\big\}},$9

The classification term is binary cross-entropy over two pruning stages, while the geometry term is a Sampson-like error. Inlier/outlier labels are generated from epipolar distances under the ground-truth essential matrix, with a match labeled an outlier if its epipolar distance exceeds $1$0. Weighted pose estimation is performed by minimizing

$1$1

Inference uses two-stage pruning with rate $1$2 per stage, followed by weighted eight-point estimation and full-size verification; with $1$3 input matches, the final set is approximately $1$4 (Xia et al., 16 Jul 2025).

5. Empirical profile of the correspondence-pruning CorrMoE

The vision CorrMoE is evaluated on in-domain, cross-scene, and cross-domain benchmarks (Xia et al., 16 Jul 2025). The in-domain datasets are YFCC100M, with 68 sequences for train/val and 4 for test, and SUN3D, with 239 sequences for train/val and 15 for test. Cross-domain evaluation uses the Zero-Shot Evaluation Benchmark (ZEB), comprising 12 datasets with diverse domains, 8 real and 4 synthetic, each with approximately 3,800 pairs across 5 overlap ratios.

The primary camera-pose metric is AUC of pose errors at $1$5, $1$6, and $1$7. Outlier-rejection metrics are Precision, Recall, and F-score, with epipolar distance threshold $1$8. On YFCC100M, CorrMoE attains pose AUC values of $1$9 at $\eCor(X,Y)=0$0, $\eCor(X,Y)=0$1 at $\eCor(X,Y)=0$2, and $\eCor(X,Y)=0$3 at $\eCor(X,Y)=0$4, compared with BCLNet’s $\eCor(X,Y)=0$5, $\eCor(X,Y)=0$6, and $\eCor(X,Y)=0$7. On SUN3D, it reports $\eCor(X,Y)=0$8, $\eCor(X,Y)=0$9, and $(Y\text{-space}, \delta_Y)$00, respectively. For YFCC100M outlier rejection, the reported scores are Precision $(Y\text{-space}, \delta_Y)$01, Recall $(Y\text{-space}, \delta_Y)$02, and F-score $(Y\text{-space}, \delta_Y)$03 (Xia et al., 16 Jul 2025).

Cross-scene splits BUC, NOT, REI, and SAC are reported in AUC@5° / F1 form as $(Y\text{-space}, \delta_Y)$04, $(Y\text{-space}, \delta_Y)$05, $(Y\text{-space}, \delta_Y)$06, and $(Y\text{-space}, \delta_Y)$07, with the text stating the strongest gain especially on SAC, at $(Y\text{-space}, \delta_Y)$08 AUC@5° over the second-best method. On the ZEB cross-domain benchmark, the average AUC@5° is reported as $(Y\text{-space}, \delta_Y)$09, with specific examples such as ETO at $(Y\text{-space}, \delta_Y)$10 and a gain of $(Y\text{-space}, \delta_Y)$11 over the second-best (Xia et al., 16 Jul 2025).

Ablation studies isolate the contributions of the dual branch and MoE fusion. On YFCC100M AUC@5°/@20°, the reported values are: Bi-Fusion MoE only, $(Y\text{-space}, \delta_Y)$12; dual-branch only, $(Y\text{-space}, \delta_Y)$13; MoE + implicit, $(Y\text{-space}, \delta_Y)$14; MoE + explicit, $(Y\text{-space}, \delta_Y)$15; full model, $(Y\text{-space}, \delta_Y)$16. The text concludes that both DDB and BF-MoE are indispensable and that the explicit branch gives stronger guidance. Depth ablation reports $(Y\text{-space}, \delta_Y)$17 as IPS $(Y\text{-space}, \delta_Y)$18, AUC@5° $(Y\text{-space}, \delta_Y)$19, AUC@20° $(Y\text{-space}, \delta_Y)$20; $(Y\text{-space}, \delta_Y)$21 as IPS $(Y\text{-space}, \delta_Y)$22, AUC@5° $(Y\text{-space}, \delta_Y)$23, AUC@20° $(Y\text{-space}, \delta_Y)$24; and $(Y\text{-space}, \delta_Y)$25 as IPS $(Y\text{-space}, \delta_Y)$26, AUC@5° $(Y\text{-space}, \delta_Y)$27, AUC@20° $(Y\text{-space}, \delta_Y)$28 (Xia et al., 16 Jul 2025).

Implementation details include PyTorch, 4 NVIDIA A100 GPUs, 500k iterations, batch size 32, Adam with initial learning rate $(Y\text{-space}, \delta_Y)$29, and decay factor $(Y\text{-space}, \delta_Y)$30 every 20k iterations. Default graph settings are $(Y\text{-space}, \delta_Y)$31 matches, KNN $(Y\text{-space}, \delta_Y)$32 in stage 1 and $(Y\text{-space}, \delta_Y)$33 in stage 2, OA clusters $(Y\text{-space}, \delta_Y)$34, and MoE configuration of 4 layers with 3 experts per layer and top-$(Y\text{-space}, \delta_Y)$35 routing (Xia et al., 16 Jul 2025).

6. CorrMoE as a correlation-aware Mixture-of-Experts for CTR prediction

In CTR prediction, CorrMoE is used as a shorthand for a correlation-aware MoE whose experts are explicitly encouraged to become de-correlated (Wang et al., 23 May 2025). The paper’s concrete implementation is called De-Correlated MoE (D-MoE). The stated motivation is that if experts learn highly correlated functions, the ensemble degenerates toward a single-expert model, gating becomes less informative, and generalization degrades. The framework therefore combines architectural de-correlation strategies with an explicit cross-expert de-correlation loss.

The architecture has three components. First, each expert owns its own embedding table:

$(Y\text{-space}, \delta_Y)$36

and for sample $(Y\text{-space}, \delta_Y)$37 the expert-specific embedding is

$(Y\text{-space}, \delta_Y)$38

Second, expert outputs are

$(Y\text{-space}, \delta_Y)$39

where experts may be homogeneous or heterogeneous, including DCNv2, xDeepFM/CIN, CrossNet, FM, and DNN. Third, the gate uses an expert-independent embedding:

$(Y\text{-space}, \delta_Y)$40

followed by aggregation

$(Y\text{-space}, \delta_Y)$41

and prediction

$(Y\text{-space}, \delta_Y)$42

The paper explicitly states that the expert-independent gating embedding is intended to avoid biasing the gate toward any expert’s private embedding space (Wang et al., 23 May 2025).

The central analytic object is Cross-Expert Correlation (CEC), defined from Pearson cross-correlation matrices over a mini-batch. For batch outputs $(Y\text{-space}, \delta_Y)$43, normalize each expert output featurewise across the batch to obtain $(Y\text{-space}, \delta_Y)$44. The cross-correlation matrix is

$(Y\text{-space}, \delta_Y)$45

and the pairwise CEC is

$(Y\text{-space}, \delta_Y)$46

The aggregate statistic is

$(Y\text{-space}, \delta_Y)$47

Lower values indicate greater de-correlation, and the metric is used both analytically and as the basis for the training penalty (Wang et al., 23 May 2025).

The de-correlation loss penalizes squared Pearson correlation:

$(Y\text{-space}, \delta_Y)$48

$(Y\text{-space}, \delta_Y)$49

The overall objective is

$(Y\text{-space}, \delta_Y)$50

with

$(Y\text{-space}, \delta_Y)$51

The paper reports that computing $(Y\text{-space}, \delta_Y)$52 on expert outputs works best compared with inputs or intermediate layers, and that Pearson-based $(Y\text{-space}, \delta_Y)$53 outperforms covariance-based variants (Wang et al., 23 May 2025).

The empirical study spans Avazu and Criteo. Avazu is reported as 40M samples with 24 categorical fields and train/valid/test splits of approximately 32.3M/4.0M/4.0M; Criteo as approximately 45M samples with 13 numerical and 26 categorical fields, using 39 fields in total after preprocessing, and train/valid/test of approximately 33.0M/8.3M/4.6M. Training uses learning rate $(Y\text{-space}, \delta_Y)$54, batch size 10,000, embedding dimension 16 for Avazu and 10 for Criteo, a gate MLP hidden size 64, and a tower MLP hidden size 500 (Wang et al., 23 May 2025).

The paper’s key design claim is the “compatibility principle”: independent embedding tables, heterogeneous experts, and the cross-expert de-correlation loss are mutually compatible and cumulative. Quantitatively, for a CIN-based two-expert MoE, CEC drops from $(Y\text{-space}, \delta_Y)$55 to $(Y\text{-space}, \delta_Y)$56 with a reported approximately $(Y\text{-space}, \delta_Y)$57 AUC gain when moving to independent embeddings. For a DCN-based MoE, replacing one homogeneous expert with a different interaction model decreases CEC from $(Y\text{-space}, \delta_Y)$58 to $(Y\text{-space}, \delta_Y)$59 and improves AUC. Adding $(Y\text{-space}, \delta_Y)$60 yields further gains; for DCNv2 on Avazu, Homo-ME baseline AUC $(Y\text{-space}, \delta_Y)$61, and for Hetero-ME (CIN+DCN), $(Y\text{-space}, \delta_Y)$62 (Wang et al., 23 May 2025).

The online result is reported on Tencent advertising platforms. Against a production Multi-Embedding MoE baseline, D-MoE/CorrMoE yields a statistically significant $(Y\text{-space}, \delta_Y)$63 GMV lift overall; across multiple scenarios, GMV lift is reported as $(Y\text{-space}, \delta_Y)$64–$(Y\text{-space}, \delta_Y)$65 and CTR lift as $(Y\text{-space}, \delta_Y)$66–$(Y\text{-space}, \delta_Y)$67. Inference cost is unchanged because the de-correlation loss is used only during training, while the dominant parameter overhead comes from expert-specific embedding tables (Wang et al., 23 May 2025).

7. Cross-cutting themes, differences, and open issues

Across the three usages, CorrMoE consistently concerns the organization of dependence or redundancy, but at different levels of abstraction. In the Earth Mover’s Correlation, the object is a discrepancy between $(Y\text{-space}, \delta_Y)$68 and $(Y\text{-space}, \delta_Y)$69 measured by $(Y\text{-space}, \delta_Y)$70 on a metric product space (Móri et al., 2020). In the CTR setting, the object is inter-expert redundancy measured by Pearson cross-correlation matrices over expert outputs (Wang et al., 23 May 2025). In the correspondence-pruning framework, the emphasis is not on a formal correlation functional but on robust feature fusion and conditional computation under cross-domain and cross-scene variability (Xia et al., 16 Jul 2025).

The computational trade-offs also differ sharply. Earth-Mover CorrMoE is broad in metric generality but expensive because empirical covariance requires an optimal transport problem from $(Y\text{-space}, \delta_Y)$71 support points to an $(Y\text{-space}, \delta_Y)$72 grid (Móri et al., 2020). The vision CorrMoE addresses scale by replacing quadratic attention with FlowAttention and by sparse top-$(Y\text{-space}, \delta_Y)$73 routing (Xia et al., 16 Jul 2025). The CTR CorrMoE adds only training-time overhead through pairwise cross-correlation computations, while leaving inference unchanged (Wang et al., 23 May 2025).

Each usage also carries its own limitations. For the Earth Mover’s Correlation, full separate-similarity invariance and the converse characterization of $(Y\text{-space}, \delta_Y)$74 remain conjectural in Banach spaces, and no closed-form asymptotic law for the empirical statistic is supplied (Móri et al., 2020). For vision CorrMoE, the reported failure cases include extreme low-texture scenes, severe viewpoint changes with very high outlier ratios, repetitive patterns, and the restriction imposed by top-$(Y\text{-space}, \delta_Y)$75 routing (Xia et al., 16 Jul 2025). For CTR CorrMoE, the method is sensitive to the de-correlation weight $(Y\text{-space}, \delta_Y)$76; if $(Y\text{-space}, \delta_Y)$77 is too large, it may force experts to differ at the expense of CTR optimization, and the gains may be limited when experts are very weak or too similar (Wang et al., 23 May 2025).

A plausible implication is that “CorrMoE” is best understood not as a single method family but as a convergent naming pattern applied to problems where correlation, correspondence, or expert redundancy is central. In current arXiv usage, the term therefore requires immediate disambiguation by domain: optimal transport dependence, visual correspondence pruning, or de-correlated expert learning.