Variate Reassembly Router in Falcon-X

• Variate Reassembly Router (VRR) is a module that reconstructs entity-specific latent trajectories from a shared prototype space for effective multivariate forecasting.
• It uses learned linear projections and a scaled dot-product soft routing mechanism to convert global latent interactions into precise variate-level outputs.
• Gated residual fusion is applied post-reassembly to balance and integrate cross-variate context with strong temporal signals, mitigating semantic interference.

Searching arXiv for the specified paper to ground the article in the cited source. The Variate Reassembly Router (VRR) is the final multivariate-specific module in Falcon-X, a time series foundation model for heterogeneous multivariate modeling. Its function is to take the globally mixed, semantically aligned latent representation produced in the shared prototype space and reconstruct it back into each entity’s original variate dimensions so that forecasting can be performed in the physical channel space. In Falcon-X, cross-variate reasoning is intentionally carried out in a dimension-agnostic latent prototype space rather than in the raw variate space; VRR is therefore the mechanism that restores entity-specific multivariate structure after latent interaction and before the quantile forecasting head (Liu et al., 26 May 2026).

1. Architectural role in Falcon-X

Falcon-X is organized around a deliberate decoupling of temporal modeling, cross-variate semantic alignment, latent interaction, and variate-space reconstruction. The pipeline proceeds as follows: raw input $\mathbf{X}$ is normalized and tokenized; a patch embedding produces $\mathbf{H}\in\mathbb{R}^{M\times P\times D}$; Time Attention yields temporal embeddings

$$\mathbf{H}^{(n)} = \mathbf{H}_T \in \mathbb{R}^{M \times P \times D};$$

Unified Prototype Diff-Attention (UPDA) maps each entity into a shared prototype space,

$$\mathbf{H}_C \in \mathbb{R}^{N \times C \times P \times D};$$

Latent Entity Attention (LEA) refines this into

$$\mathbf{H}_C' \in \mathbb{R}^{N \times C \times P \times D};$$

VRR reconstructs entity-specific latent trajectories,

$\mathbf{H}_V \in \mathbb{R}^{M\times P\times D};$

and finally $\mathbf{H}_V$ is fused with $\mathbf{H}_T$ and passed to the quantile forecasting head (Liu et al., 26 May 2026).

This positioning makes VRR the bridge from global latent interaction back to per-variate outputs. The reassembly step is necessary because Falcon-X performs cross-variate reasoning in a space of fixed prototype dimension $C$, whereas real inputs have heterogeneous variate counts $m_i$ per entity. Without such a mechanism, the model would have no principled way to recover predictions aligned to the original physical channels. The paper further emphasizes that directly forcing global context into raw-space channels can introduce semantic interference, especially when inter-variate dependencies are weak; VRR, particularly together with gated residual fusion, is designed to mitigate that effect (Liu et al., 26 May 2026).

2. Latent interfaces, inputs, and outputs

For each entity $\mathbf{H}\in\mathbb{R}^{M\times P\times D}$0, VRR consumes three learned projections derived from the temporal and prototype-refined features:

$\mathbf{H}\in\mathbb{R}^{M\times P\times D}$1

Here $\mathbf{H}\in\mathbb{R}^{M\times P\times D}$2 denotes the temporal embedding of entity $\mathbf{H}\in\mathbb{R}^{M\times P\times D}$3 after Time Attention, and $\mathbf{H}\in\mathbb{R}^{M\times P\times D}$4 denotes the refined prototype-space representation after LEA. The router returns a reconstructed variate-specific representation

$\mathbf{H}\in\mathbb{R}^{M\times P\times D}$5

which over all entities forms

$\mathbf{H}\in\mathbb{R}^{M\times P\times D}$6

Tensor	Source	Role
$\mathbf{H}\in\mathbb{R}^{M\times P\times D}$7	$\mathbf{H}\in\mathbb{R}^{M\times P\times D}$8	Routing Request
$\mathbf{H}\in\mathbb{R}^{M\times P\times D}$9	$$\mathbf{H}^{(n)} = \mathbf{H}_T \in \mathbb{R}^{M \times P \times D};$$0	Prototype Index
$$\mathbf{H}^{(n)} = \mathbf{H}_T \in \mathbb{R}^{M \times P \times D};$$1	$$\mathbf{H}^{(n)} = \mathbf{H}_T \in \mathbb{R}^{M \times P \times D};$$2	Source Context

The paper defines the Routing Request as a query-like entity identity tag encoding the original variate’s trajectory and dimensional identity. The Prototype Index is an addressable map used to match the request to the latent prototype library. The Source Context is the semantic payload that will be retrieved and reassembled. This factorization separates addressing from content: the router can both locate the relevant latent semantic slots and retrieve the content required to reconstruct the variate-specific output (Liu et al., 26 May 2026).

3. Request-and-dispatch routing mechanism

VRR is described as a request-and-dispatch router. The original temporal trajectory supplies a request signal, and the prototype-space state provides the indexed latent content to be dispatched back into the original variate layout. The request is first formed from the temporal features,

$$\mathbf{H}^{(n)} = \mathbf{H}_T \in \mathbb{R}^{M \times P \times D};$$3

while the prototype index and source context are constructed from the prototype-refined latent state,

$$\mathbf{H}^{(n)} = \mathbf{H}_T \in \mathbb{R}^{M \times P \times D};$$4

The core routing equation is a scaled dot-product soft-routing rule:

$$\mathbf{H}^{(n)} = \mathbf{H}_T \in \mathbb{R}^{M \times P \times D};$$5

This computation produces soft assignment weights between the request and the prototype index, then uses those weights to dispatch the source context into reconstructed variate-level features. The result is a representation

$$\mathbf{H}^{(n)} = \mathbf{H}_T \in \mathbb{R}^{M \times P \times D};$$6

which restores the original physical dimensionality for entity $$\mathbf{H}^{(n)} = \mathbf{H}_T \in \mathbb{R}^{M \times P \times D};$$7 and contributes to the full reconstructed tensor $$\mathbf{H}^{(n)} = \mathbf{H}_T \in \mathbb{R}^{M \times P \times D};$$8 (Liu et al., 26 May 2026).

A central property of this design is that reconstruction is not performed by collapsing everything into a generic shared representation. Instead, the latent context is routed back according to entity-specific requests derived from the temporal backbone. This suggests that variate identity is preserved through reconstruction even though interaction itself occurs in the shared prototype space.

4. Gated residual fusion and preservation of temporal signal

After reconstruction, Falcon-X does not directly replace temporal features with the routed latent context. Instead, it applies a gated residual fusion:

$$\mathbf{H}^{(n)} = \mathbf{H}_T \in \mathbb{R}^{M \times P \times D};$$9

In this expression, $$\mathbf{H}_C \in \mathbb{R}^{N \times C \times P \times D};$$0 is a linear layer followed by sigmoid, $$\mathbf{H}_C \in \mathbb{R}^{N \times C \times P \times D};$$1 is element-wise multiplication, $$\mathbf{H}_C \in \mathbb{R}^{N \times C \times P \times D};$$2 preserves the strong per-variate temporal signal, and $$\mathbf{H}_C \in \mathbb{R}^{N \times C \times P \times D};$$3 injects cross-variate context selectively (Liu et al., 26 May 2026).

The stated purpose of this gate is to avoid over-injecting cross-variate information into weakly correlated datasets. The model therefore filters routed global context depending on the current input rather than applying it indiscriminately. The paper explicitly connects this mechanism to robustness against semantic interference when inter-variate dependencies are weak. In that sense, VRR is not only a reconstruction device but part of a controlled fusion interface between latent multivariate reasoning and the temporal backbone.

The significance of this choice becomes clearer in relation to Falcon-X’s architectural premise. Because cross-variate interaction happens in a semantically aligned latent space rather than in raw channel coordinates, the final step must both recover physical-channel structure and preserve local temporal fidelity. The gated residual pathway is the component that enforces that balance.

5. Dependence on UPDA and LEA

VRR depends directly on the outputs of the two latent-space interaction modules that precede it. UPDA is responsible for projecting heterogeneous raw variates into a shared prototype space,

$$\mathbf{H}_C \in \mathbb{R}^{N \times C \times P \times D};$$4

UPDA itself uses dual prototype keys

$$\mathbf{H}_C \in \mathbb{R}^{N \times C \times P \times D};$$5

and constructs positive and negative affinities. The resulting prototype representation is

$$\mathbf{H}_C \in \mathbb{R}^{N \times C \times P \times D};$$6

VRR does not directly use these equations, but it relies on the aligned latent prototype representation created by UPDA to make reassembly meaningful (Liu et al., 26 May 2026).

LEA then refines the prototype-space representation by mixing information globally across entities and prototypes:

$$\mathbf{H}_C \in \mathbb{R}^{N \times C \times P \times D};$$7

VRR consumes $$\mathbf{H}_C \in \mathbb{R}^{N \times C \times P \times D};$$8 rather than the raw $$\mathbf{H}_C \in \mathbb{R}^{N \times C \times P \times D};$$9, so its routing input already contains globally contextualized information. In the paper’s own decomposition, UPDA aligns heterogeneous variates into a shared latent coordinate system, LEA allows all aligned latent entities to interact, and VRR routes the resulting global context back into the original variate layout (Liu et al., 26 May 2026).

This modular dependency clarifies the division of labor inside Falcon-X. UPDA provides semantic alignment across heterogeneous physical quantities; LEA performs cross-variate interaction in the shared space; VRR reconstructs entity-specific trajectories in the physical channel space. A plausible implication is that Falcon-X’s multivariate capability depends not on any one module in isolation but on the compatibility between latent alignment, latent interaction, and variate-space reassembly.

6. Training, implementation, and empirical interpretation

The paper does not define a separate loss exclusively for VRR. Instead, the router is trained end-to-end through the overall forecasting objective,

$$\mathbf{H}_C' \in \mathbb{R}^{N \times C \times P \times D};$$0

where $$\mathbf{H}_C' \in \mathbb{R}^{N \times C \times P \times D};$$1 is the quantile loss on forecast outputs and $$\mathbf{H}_C' \in \mathbb{R}^{N \times C \times P \times D};$$2 is the orthogonality loss for prototype separation,

$$\mathbf{H}_C' \in \mathbb{R}^{N \times C \times P \times D};$$3

Accordingly, VRR is supervised indirectly: its utility is measured by whether it improves the final quantile forecasts rather than by a dedicated reconstruction criterion (Liu et al., 26 May 2026).

Several implementation choices are specified. Parallel entity-aware masking is used so that routing can be done concurrently across all entities without explicit loops. The router uses independent linear projections to produce request, index, and context tensors. It employs scaled dot-product soft routing rather than hard assignment. The design is intentionally lightweight compared with dense variate-level attention, and because routing operates from a fixed prototype space, the method avoids dependence on the raw variate count (Liu et al., 26 May 2026).

The ablation study does not isolate VRR in a standalone numeric row, but the paper reports evidence for the importance of the routing and fusion design. Removing the gated residual connection degrades performance. The inference analysis shows that multivariate inference improves accuracy over univariate inference mode while preserving univariate performance. The case studies show that the model handles strongly positively correlated channels, negatively correlated channels, and weakly related settings without overfitting the cross-variate context. The paper’s qualitative conclusion is that the routing/reassembly stage allows Falcon-X to benefit from cross-variate information while avoiding semantic collapse (Liu et al., 26 May 2026).

An important clarification follows from these observations. VRR is not presented as a generic attention block operating in raw channel space; it is specifically a prototype-to-variates reconstruction module embedded in an architecture that performs cross-variate modeling in a shared prototype space. Its purpose is therefore narrower and more structural: to convert globally contextualized latent information back into entity-specific multivariate representations through request-and-dispatch soft routing, then fuse those representations with temporal features before probabilistic forecasting.