RadarGate: LoRA MoEs & mmWave Spy Radar

• RadarGate in LLMs augments traditional MoE gating by applying learnable rotational transformations to LoRA outputs, thereby expanding the hypothesis space and enhancing model generalization.
• RadarGate as a radar system uses commercial mmWave hardware combined with FFT analysis, CNN classification, and MUSIC-based AoA estimation to achieve high-accuracy detection and localization of covert radars.
• Empirical results show that RadarGate improves task accuracy, convergence stability, and low-shot performance in LLMs while delivering robust, real-time performance with minimal false alarms in mmWave environments.

RadarGate refers to two distinct innovations in recent research literature: (1) a geometrically-inspired gating mechanism for scaling Low-Rank Adaptation (LoRA)-based Mixture-of-Experts (MoE) architectures in LLMs, and (2) a passive detection and localization system for identifying covert mmWave radars using commercial off-the-shelf (COTS) mmWave hardware. The following article delineates the technical principles, system architectures, and empirical results for both usages, as defined in the literature (Guo et al., 29 May 2025, Qiu et al., 2022).

1. Geometric Gating for Scalable LoRA-Based MoEs in LLMs

RadarGate, in the context of neural architectures, denotes a gating mechanism aimed at mitigating underfitting and poor generalization in large-scale LoRA-MoEs. Traditional MoE gating maps each input to a convex combination of expert outputs, computed via a softmax-weighted sum over LoRA adapter representations. This induces two key limitations: underfitting (restricted by the function class of simple gates) and poor expressivity (the output is confined to the convex cone generated by LoRA outputs).

RadarGate Mechanism

RadarGate augments the gating process by introducing learnable, input-dependent rotational transformations on each LoRA's output, informed by cross-LoRA feature interactions. The method comprises the following components:

• LoRA Representation Construction: For each input $x$, the $i$-th adapter output is $v_i(x) = xA_iB_i$, and a reference sum $s_i(x) = \sum_{j \neq i} v_j(x)$ is computed.
• Fusion via Elementwise Hadamard Product: A fused vector $u_i(x) = v_i(x) \odot s_i(x)$ captures joint information.
• Rotational Gate (RotationGate): Raw rotation angles $\alpha_{r_i}(x)$ are generated via a linear transformation with learnable matrix $\theta_r$. A block-diagonal rotation matrix $R_i(\alpha_{r_i})$ is constructed, and the LoRA output is rotated to yield $\tilde{v}_i(x) = R_i(\alpha_{r_i}(x)) v_i(x)$.
• Magnitude Gate (StretchGate): Standard top-$k$ softmax gating $g(x)$ is applied over normalized projections of the rotated features.
• Final Layer Aggregation: The layer output is $y(x) = xW + \sum_{i=1}^n g_i(x)\tilde{v}_i(x)$, where $W$ is the backbone weight matrix.

2. Mathematical Properties and Expressiveness

RadarGate expands the hypothesis space and output region of MoEs:

• Convex Cone Expansion: In conventional gating, outputs are constrained to $H(x) = \mathrm{conv}\{v_i(x)\}$. RadarGate’s rotations allow $H'(x) = \mathrm{conv}\{\tilde{v}_i(x)\}$, which strictly contains $H(x)$, conferring additional representational flexibility.
• Function Class Enlargement: The composite gating function class incorporating both magnitude and rotation parameters strictly contains the ordinary magnitude-only gating class, guaranteeing non-increasing training error and improved fit for complex adaptation targets.
• Feature Interaction: The fusion and rotation steps induce direct coupling across LoRA adapters, enabling the model to discover and align semantically correlated representations and decorrelate unrelated ones.

3. Contrastive Geometric Effects

Empirical investigations reveal that RadarGate’s learned rotations are contrastive:

• Alignment for Semantically Similar Adapters: When two LoRA adapters encode similar functionality, their post-rotation outputs become closely aligned (cosine similarity increases), minimizing the relative angle.
• Separation for Dissimilar Adapters: Irrelevant or unrelated adapters are rotated to become more orthogonal, enforcing a margin and aiding discriminability.

This geometric alignment is quantifiable by tracking cosine similarity between rotated vectors and is realized progressively over training.

4. Experimental Protocols and Technical Benchmarks

Experiments assess RadarGate across 21 tasks and six public benchmarks (GLUE, MMLU, WMT14, GPQA, MATH, GSM8K). Metrics include task-specific accuracy and generalization performance. Baseline comparisons span both rule-based (LoRaHub, Arrow, PEMs) and learnable gating approaches (HydraLoRA, MoLE, OMoE, Nexus, Tutel). RadarGate consistently achieves:

• Superior Fitting: Matches or exceeds all baselines in >90% of tasks, with improvements up to +20 percentage points.
• Superior Generalization: Gains of +30–50% over rule-based methods and +5–10% over learnable baselines in zero-shot routing.
• Monotonic Module Scaling: Accuracy improves steadily as the number of LoRA modules increases (n=5→40), with up to +8 percentage point gain, in contrast to U-shaped or degrading performance in baselines.
• Rapid, Stable Convergence: Training loss curves show reduced oscillations and faster convergence.
• Robust Low-shot Learning: Outperforms all baselines by >5 percentage points on average with only 50 examples.
• Model Scaling: Maintains a +5–10% edge as model size scales from 110M to 8B parameters (Guo et al., 29 May 2025).

5. Passive mmWave Spy Radar Detection and Localization

In a separate technical context, RadarGate refers to a practical system for identifying and localizing unauthorized mmWave radars—termed spy radars—using a single passive COTS mmWave radar device (Qiu et al., 2022). The approach is divided into detection, classification, and localization modules as follows:

• Frequency-Component Detection: Implements single-tone and linear-chirp demodulation to detect any mmWave transmission in the 60–80 GHz range. Detection metrics are computed from the FFT power spectrum via ratio and absolute threshold tests, with the noise baseline estimated by the statistical median.
• Waveform Classification: A CNN operates on normalized FFT magnitude vectors ($1 \times 1024$) to distinguish radar signals from other mmWave emissions (e.g., WiGig IEEE 802.11ad/ay). The classifier achieves >99% accuracy on held-out data.
• Localization (MUSIC AoA + Triangulation): The system estimates Angle-of-Arrival (AoA) at multiple anchor points using MUSIC spatial spectrum estimation. The 2D location is computed via least-squares intersection of measured AoA rays. When multiple radars are present, multi-peak AoA detection, association, and clustering algorithms enable simultaneous localization.

Empirical results indicate:

• Spy radar detection rate ≥96% (100% for indoor scenarios up to 20 m).
• False alarm rates between 2.6–4.1%.
• Cross-validated classification accuracy >99%.
• 2-anchor least-squares localization achieves ≤0.256 m error at the 90th percentile, improving to ≈0.15 m with 6 anchors.
• Robustness against multipath, human blockage, and environmental variation.

In summary, this RadarGate system operationalizes efficient, passive, and precise countermeasures for privacy-sensitive mmWave environments.

6. Broader Significance and Applications

The RadarGate framework, in both neural architecture design and electromagnetic surveillance, addresses pressing scalability and security challenges:

• In LLM adaptation, RadarGate permits modular, scalable, and generalizable adaptation across diverse tasks, overcoming long-standing bottlenecks in the convexity and restricted hypothesis space of gated MoEs.
• In security, the RadarGate system offers an effective, deployable tool for privacy assurance in spaces vulnerable to covert mmWave surveillance, leveraging standard commercial hardware and robust signal processing pipelines.

RadarGate thus denotes state-of-the-art solutions in two domains, unified by principles of modular signal transformation and selective, learned gating. Each variant is independently validated on benchmarks appropriate to its domain, with technical rigor and reproducibility prioritized by the respective research efforts (Guo et al., 29 May 2025, Qiu et al., 2022).