SLM-MUX: Multidomain Multiplexing

• SLM-MUX is a class of multiplexing architectures that selectively processes spatial, modal, electrical, or model-based channels across diverse domains.
• Implementations span spatial mode conversion with low crosstalk in fiber optics, on-chip mode and wavelength switching in integrated photonics, and superconducting circuits for cryogenic quantum systems.
• SLM-MUX enhances system performance with optimized metrics—such as crosstalk < -23 dB and SNR improvements up to 3 dB—while enabling adaptable orchestration in machine learning models.

SLM-MUX refers to a class of multiplexing architectures and devices in photonics, microwave electronics, and machine learning, where multiple independent channels—spatial, modal, electrical, or model-based—are selectively processed, routed, or orchestrated for high-efficiency, low-loss, or improved accuracy. Implementations span multiplexers for spatial modes in fiber optics, reconfigurable switches for integrated photonics, superconducting demultiplexers in cryogenic electronics, and multi-model orchestration frameworks in natural language processing. The following sections give a technical analysis of SLM-MUX across its various domains.

1. Physical SLM-MUX: Spatial Mode Multiplexing and Multi-Plane Light Conversion

The term “SLM-MUX” in physical-layer optical systems most commonly denotes spatial mode multiplexers that exploit spatial light modulation techniques for channel-selective transformation and routing. A notable realization leverages Multi-Plane Light Conversion (MPLC), which performs unitary transformations between input Gaussian modes and orthogonal fiber eigenmodes via a cascaded sequence of phase-only masks separated by Fourier-like free-space propagation (Labroille et al., 2014). The transformation,

$$U |\psi_\text{in}\rangle = |\psi_\text{out}\rangle,$$

is lossless in the ideal case, as each phase mask imparts only a spatial phase profile. The modal overlap metric,

$$\rho_{xx} = \left| \int w_p^*(\vec{r})\, w_{LP_{xx}}(\vec{r})\, d\vec{r} \right|^2,$$

quantifies conversion fidelity. Reported MPLC SLM-MUX devices achieve mode selectivity better than $-23\ \text{dB}$ (low crosstalk), intrinsic conversion loss below $1.2\ \text{dB}$ (non-idealities only), and total insertion loss of $-4.1\ \text{dB}$ ($-2.4\ \text{dB}$ with improved dielectric coatings), with performance constant across the C-band. Arbitrary mode conversions are achievable by tailoring phase profiles, demonstrated for six orthogonal LP modes including $LP_{01}$, $LP_{11a}$, $LP_{11b}$, etc.

This approach enables loss-efficient, broadband spatial multiplexing for few-mode fiber systems; further optimization of optical coatings, cavity design, and phase-mask engineering can push device performance toward the theoretical minimum loss.

2. Integrated Multiplexers: On-Chip Mode and Wavelength Switching

SLM-MUX architectures have been adapted to photonic integrated circuits for simultaneous mode-division multiplexing (MDM) and wavelength-division multiplexing (WDM) (Stern et al., 2015). The canonical design splits processing into: (1) independent mode conversion via phase-matched ring resonators, (2) single-mode switching backbone with thermo-optic microring switches, and (3) mode reconversion for output. Such a convert–process–reconvert sequence decouples modal differences and allows identical switching circuits for all channels.

Performance metrics include intermodal crosstalk $< -20$ dB, bit-error rates below $10^{-9}$, power penalties under $2.4$ dB for simultaneous multiplexing, overall insertion loss between $5.4$ and $9.1$ dB, and tunable bandwidth per switching element (~16 GHz/ring). Recent prototypes demonstrate flexible, dynamically configurable multiplexing for chip-scale networks, data centers, and emerging on-chip interconnect paradigms.

Key switching mechanism:

Stage	Functionality	Key Component
Conversion	Multimode $\to$ single-mode transformation	Ring resonators, tapers
Switching	Signal routing, WDM/MDM selection	Microring heaters
Reconversion	Single-mode $\to$ multimode restoration	Mode reconverters

This integration advances the scalability and granularity of photonic multiplex networks, achieving high channel density within compact footprints.

3. SLM-MUX in Cryogenic and Quantum Electronics

SLM-MUX devices in cryogenic quantum systems comprise scalable electrical multiplexing circuits that route signals without significant heat dissipation or loss. Implementations include Wheatstone bridges of tunable SQUID inductors (“Tunable Inductor Bridge,” TIB) enabling transmit, reflect, and invert modes (Chapman et al., 2016). Operation is governed by flux-tunable arms:

$$T = \frac{i\omega(l_1 - l_2) Z_0}{(i\omega l_1 + Z_0)(i\omega l_2 + Z_0)}$$

Switching speed $\leq 15$ ns, on-off ratios $>20$ dB (up to $40$ dB near $6$ GHz), insertion loss $< 0.5$ dB, and readout error rates $<0.01$ for binary signaling are demonstrated. The TIB facilitates code-domain multiplexing (via Walsh code modulation) and time-domain multiplexing, with direct utility in superconducting qubit array readout, as well as broader cryogenic multiplexing applications.

Solid-state variants employ cascaded JoFET switches (InAsOI/Al/HfO2 stacks), achieving hierarchical $1$-to-$8$ multiplexing with ON/OFF ratios of $17.5$ dB, near-zero insertion loss, and MHz–GHz operation at $T=50$ mK (Paghi et al., 15 Oct 2024). Selective gating and optimized routing mitigate parasitic capacitance effects, opening the path for massive reduction of cryostat I/O complexity in scalable quantum electronics.

4. SLM-MUX for Lattice Modulation in Communications

In communications, SLM-MUX also describes “Spatial Lattice Modulation” for MIMO systems, where channel bits are directly mapped onto high-dimensional lattice points across spatial, in-phase, and quadrature dimensions (Choi et al., 2018). Instead of choosing active antennas and symbols separately, the method constructs a joint constellation in $\mathbb{R}^{2N_t}$ via

$\{0\} \cup \mathcal{C}$

per dimension, with the “0” symbol indicating antenna/component inactivity. Entropy per channel increases to $2N_t \log_2(M+1)$. Dense lattices (e.g., Barnes–Wall) boost minimum Euclidean distance $d_{\min}$ and coding gain $\gamma_c$, yielding up to $1.5$–$3$ dB SNR improvements and reduced average symbol vector error probability (ASVEP). Efficient detection is supported by lattice sphere decoding; complexity drops from $O((M+1)^{2N_t})$ to $O(N_t^3)$.

This paradigm provides increased spectral efficiency and robust noise immunity relative to conventional SM and spatial multiplexing approaches.

5. Reconfigurable SLM-MUX for High-Throughput Characterization

SLM-MUX architectures for electrical measurement multiplexing at cryogenic temperatures use multi-level selective gating (MLSG) schemes (Bian et al., 27 Mar 2024). By hierarchically branching channels via addressable gates, $K$ control lines yield $2^K$ output channels per MUX chip, scaling exponentially. A prototype with four chips achieves $128$ interconnects using just $14$ wires from room temperature; leakage currents are suppressed to $<1$ pA (off-resistance $\sim\!30$ TΩ/channel), supporting high-throughput device screening—demonstrated for $128$ graphene nanogap quantum devices within a single cooling cycle.

Element	Output Multiplex Ratio	Control Lines Needed
5-level MUX	$32$	$10$
$N$ MUX chips	$N \times 32$	$N$ signal wires + $10$

This design optimizes wiring resources and interconnect flexibility for systematic quantum device characterization, facilitating broader experimental capabilities in quantum transport and related domains.

6. Model-Orchestrating SLM-MUX in Language Processing

SLM-MUX extends to a distinct domain in machine learning as an orchestration architecture for small LLMs (SLMs) (Wang et al., 6 Oct 2025). The framework operates in two phases:

1. Independent Generation: Each SLM independently samples $k$ candidate answers at nonzero temperature ($T = 0.3$), producing frequencies $f_i(y)$ for answer $y$.
2. Confidence Estimation and Selection: Each model’s most frequent answer is selected with corresponding confidence; ties are resolved by historical validation accuracy. Answers are “muxed” without inter-model dialogue.

Optimization strategies include:

• Model Selection Search: Selecting model combinations that maximize union accuracy minus contradiction penalty, with exhaustive search given small candidate pools.
• Test-Time Scaling: Balancing number of models and samples per model to optimize accuracy, observing diminishing returns with excess diversity or insufficient confidence disambiguation.

SLM-MUX’s confidence-driven output selection yields up to $13.4\%$ improvement on MATH, $8.8\%$ on GPQA, $7.0\%$ on GSM8K over prior orchestration and individual model approaches, with ensembles of two SLMs matching or exceeding Qwen2.5 72B on targeted reasoning benchmarks. Theoretical analysis supports accuracy improvement for models with per-sample accuracy $p > 0.5$:

$$A(N, p) = \text{tail probability of Binomial}(N, p)$$

SLM-MUX avoids error amplification associated with debate-based orchestration, and can be adapted for cost-efficient or edge deployment scenarios where model size and compute are constrained.

7. Comparative Overview and Technological Impact

Across implementations, SLM-MUX architectures:

• Enable high-fidelity, low-loss channel multiplexing (optics, electronics, quantum).
• Achieve scalable chip-level integration, minimizing physical and energy overhead in dense environments.
• Provide robust error and crosstalk suppression (e.g., mode selectivity $<-23$ dB, error $<10^{-2}$).
• Expand spectral, modal, and functional throughput via mode-selective, configurable, and orchestration mechanisms.
• In machine learning, enable practical accuracy improvements for ensembles of efficient models, rivalling larger alternatives.

The convergence of these principles in SLM-MUX advances bandwidth, scaling, multiplexing efficiency, and adaptive orchestration across photonics, quantum electronics, communications, and intelligent systems. Future research anticipates dynamic orchestration, enhanced calibration methods, further loss reduction, and integration with reconfigurable and hybrid multiplexed environments.