Recursive Connector in Multimodal & Quantum Systems

• Recursive connector is a specialized module that enables iterative information transfer and feature fusion across repeated processing stages in complex systems.
• In multimodal transformers, it refines intermediate vision and text representations using RMS normalization and modality-specific MLPs to maintain consistent scaling.
• In quantum certification, it implements local linear maps to coarse-grain multipartite systems while preserving critical quantum properties under strict constraints.

A recursive connector is a specialized computational or algebraic module designed to enable information transfer, alignment, or coarse-graining across repeated (recursive) processing stages in complex systems. The term appears prominently in two distinct research domains: large multimodal model architectures that utilize transformer decoders with iterative refinements (Xu et al., 9 Feb 2026), and tensor network approaches to scalable quantum certification via local coarse-graining transformations (Navascues et al., 2019). In each context, the recursive connector serves as the key mechanism by which intermediate representations or system components are fused, projected, or contracted to enable on-demand refinement, scale-bridging, or property-preserving simplifications.

1. Recursive Connector in Multimodal Transformers

In the context of large multimodal models (LMMs), the recursive connector is a network module introduced in the RecursiveVLM architecture to align features and re-inject fused information across recursion steps within a shared-parameter transformer decoder. At every recursion step $r$, after propagating an input multimodal embedding $E^{(r)} = [V^{(r)}; T^{(r)}]$ through the $L$-layer backbone, the recursive connector does not simply pass the deepest hidden states to the next iteration. Instead, it performs the following operations:

• Samples a subset of layers $S = \{\ell_1, ..., \ell_k\}$ (commonly four uniformly spaced layers within $[1, L]$).
• Decomposes each selected $H_\ell^{(r)} \in \mathbb{R}^{N \times d}$ into vision ($V_\ell^{(r)}$) and text ($T_\ell^{(r)}$) blocks to respect modality distinctions.
• Applies distinct connector MLPs $C_{\ell,v}, C_{\ell,t}$ for each modality, consisting of RMS normalization, a modality-specific MLP with up- and down-projection, and a learnable per-dimension residual scale.
• Sums the corrections for each modality on top of the original embeddings $V^{(1)}$, $T^{(1)}$, producing the input for the next recursion: $E^{(r+1)}$.

This approach ensures that each recursion operates on feature inputs of consistent scale and leverages fused representations from multiple intermediate depths (Xu et al., 9 Feb 2026).

2. Mathematical Structure and Modality-Specific Projections

Mathematically, for each recursion $r$ and selected layer $\ell \in S$: $$\begin{align*} H_\ell^{(r)} &= [V_\ell^{(r)} ; T_\ell^{(r)}] \ \tilde{V}_\ell^{(r)} &= \text{RMSNorm}(V_\ell^{(r)}) \ \tilde{T}_\ell^{(r)} &= \text{RMSNorm}(T_\ell^{(r)}) \end{align*}$$ The connector MLPs operate as: $$\begin{align*} A_{\ell,v}^{(r)} &= \tilde{V}_\ell^{(r)} \odot s_{\ell,v} + \sigma (\tilde{V}_\ell^{(r)} W_{\ell,v}^u) W_{\ell,v}^d \ A_{\ell,t}^{(r)} &= \tilde{T}_\ell^{(r)} \odot s_{\ell,t} + \sigma (\tilde{T}_\ell^{(r)} W_{\ell,t}^u) W_{\ell,t}^d \end{align*}$$ where $s_{\ell,v}, s_{\ell,t} \in \mathbb{R}^d$ are learnable scales, and $W^u, W^d$ are up/down-projection matrices. The next-step embeddings are: $$\begin{align*} V^{(r+1)} &= V^{(1)} + \sum_{\ell \in S} A_{\ell,v}^{(r)} \ T^{(r+1)} &= T^{(1)} + \sum_{\ell \in S} A_{\ell,t}^{(r)} \ E^{(r+1)} &= [V^{(r+1)} ; T^{(r+1)}] \end{align*}$$ Vision and text modalities are projected with independent parameter sets. This is essential to accommodate distributional and statistical differences (e.g., vision features typically have differing norms and dispersions relative to language tokens), thus preventing modality misalignment (Xu et al., 9 Feb 2026).

3. Alignment Across Recursion Steps and Monotonicity Guarantees

To ensure stability and effectiveness across recursion depths, RMS normalization is used to equalize input norms between recursion steps and prevent scale drift. The recursive connector parameters are zero-initialized, ensuring that at $r=1$, the model reproduces standard pretraining behavior ($E^{(2)} = E^{(1)}$), which stabilizes downstream training. Critically, the RecursiveVLM employs a Monotonic Recursion Loss, supervising the output at each recursion. If the cross-entropy loss for any token increases at a step, it is upweighted by a factor $\beta > 1$, and the total training loss aggregates all recursion steps. The tight alignment enforced by the connector increases the probability that each recursion step either reduces or maintains per-token loss, enforcing monotonic improvement (Xu et al., 9 Feb 2026).

4. Recursive Connector Tensor Networks for Quantum Certification

In quantum information theory, especially for scalable quantum certification, recursive connectors are defined as local linear maps $$\Omega: V_{S_1 \otimes \dots \otimes S_m} \to V_{T_1 \otimes \dots \otimes T_q}$$, where the domain and codomain are vector spaces associated with multipartite quantum or generalized probabilistic systems. These connectors are repeated recursively across system layers to coarse-grain an $N$-site system into smaller blocks while preserving crucial properties such as Bell nonlocality, separability, or quantum realizability (Navascues et al., 2019).

The connector must satisfy the "no-rescaling-hardening" (NRH) condition, which guarantees that, for any extension system $T$,

$$(\Omega \otimes \text{id}_T)(v_{in}) \in C_{T_1 \dots T_q \otimes T}$$

and that normalization is not exceeded: $$e_{T_1 \dots T_q} \otimes e_T \left( (\Omega \otimes \text{id}_T)(v_{in}) \right) \leq e_{S_1 \dots S_m} \otimes e_T(v_{in})$$ This structure allows for recursive contraction, yielding a top-level witness functional $W(P)$ with an explicit tensor-network form using only local connectors.

5. Recursive Application, Constraints, and Witness Extraction

Recursive application involves selecting blocks (e.g., $m$ adjacent sites) and applying the same connector across all blocks. After $L$ recursive coarse-graining steps, the original $N$-site system is reduced to a tractable size $s$ suitable for direct witness evaluation. Specific constraints are imposed on the recursive connectors:

• For Bell locality: Each $m \to 1$ connector must preserve local deterministic structures.
• For separability: Each $\Omega$ must map fully separable states to separable ones, which can be formulated as LP or SDP constraints.
• For quantum realizability: Each connector must map valid quantum states to quantum states, verifiable within the NPA hierarchy using SDP tests (Navascues et al., 2019).

The final witness value, if it verifies violation or nonclassicality in the small system, constitutes a certificate for the property in the full $N$-site system, with the explicit witness decomposable into the recursive connector tensor network.

6. Computational Properties and Implementation Considerations

The computational complexity of each coarse-graining recursion layer is $O(N d^m)$, with overall scaling $O(N d^m \log N)$ for fixed block size $m$, which is linear in $N$ up to logarithmic factors. The methodology extends to systems with MPS or PEPS representations, where block contractions remain efficient for moderate bond dimensions $\chi$ (Navascues et al., 2019). In the multimodal transformers setting, connector parameters remain lightweight due to per-modality partitioning and re-use across recursion steps, and the primary backbone parameters are shared across depth, preventing growth in total parameter count (Xu et al., 9 Feb 2026).

7. Comparative Summary

Application Domain	Recursive Connector Role	Key Mechanism
Multimodal Transformers (RecursiveVLM) (Xu et al., 9 Feb 2026)	Fuses and aligns intermediate representations across recursion steps; modality-specific refinement	RMSNorm, MLPs, modality-specific projections, additive correction
Quantum Certification (Tensor Networks) (Navascues et al., 2019)	Coarse-grains multipartite systems, preserving nonclassical properties for scalable certification	Local linear maps (tensors) with property-preserving constraints, NRH condition

Recursive connectors provide a principled mechanism for iterative signal refinement in deep learning and recursive coarse-graining in tensor networks. Their formal design—whether rooted in distributional symmetry for multimodal embeddings or in cone-preserving linearity for quantum systems—enables scalable, property-preserving computation in high-dimensional and recursive architectures.